Collected public posts from the past 30 days, newest first, with original sources. Scores belong to each post's date.
Puget disputes a description of LLMs as only matrix multiplications, emphasizing the role of nonlinear operations. Commentary on a quoted consciousness argument; no ARC result claimed.
General AI commentary from a watched contender; background colour only.
Posted .
Isuru Lakmal shared a claimed rank of 1713 and linked the ARC Prize 2026 ARC-AGI-3 competition. The link resolves to the contest landing page, not a team or roster.
Low-priority participant discovery lead only; exact Kaggle identity, team ID and score remain unresolved. No medal or score-gain claim.
Posted .
Linked releaseMazur replies with a rainbow emoji to @BinxNet praising his AI-generated vistas gallery.
Routine organizer capture; light community colour, no benchmark or competition result.
Posted .
Linked releaseItsu replies to their ARC-AGI-2 Kaggle-share post: “As I was saying that, now sixth.” The root says ゴイサギ (a bird pun) and links the ARC-AGI-2 contest.
Fresh attributed standings reaction; this is not an independently verified leaderboard rank, score, medal or roster change.
Posted .
In a reply to Ben Sharpe suggesting replacing the Roku screensaver with his own vistas, Mazur jokes that it is on his mind but there is not enough time in the day.
Organizer capture before selection: light personal-project banter; no ARC result.
Posted .
Linked releaseReplying to a question about his arXiv tooling, Mazur says one of his projects is an AI research assistant for arXiv and links Emergent Mind.
Organizer capture: general AI project context; no ARC score, contest milestone or new benchmark result.
Posted .
Linked releaseFourth media-only follow-up with animated examples in Mazur’s vista-gallery thread. No accompanying text or benchmark claim.
Organizer capture: general AI project context; no ARC score, contest milestone or new benchmark result.
Posted .
Linked releaseThird media-only follow-up with animated examples in Mazur’s vista-gallery thread. No accompanying text or benchmark claim.
Organizer capture: general AI project context; no ARC score, contest milestone or new benchmark result.
Posted .
Linked releaseSecond media-only follow-up with animated examples in Mazur’s vista-gallery thread. No accompanying text or benchmark claim.
Organizer capture: general AI project context; no ARC score, contest milestone or new benchmark result.
Posted .
Linked releaseFirst media-only follow-up with animated examples in Mazur’s vista-gallery thread. No accompanying text or benchmark claim.
Organizer capture: general AI project context; no ARC score, contest milestone or new benchmark result.
Posted .
Linked releaseMazur shares a gallery of 50 animated vistas he says were created with Opus 5.5, inspired by a Roku screensaver. He describes slow scrolling scenes and changing art styles.
Organizer capture: general AI project context; no ARC score, contest milestone or new benchmark result.
Posted .
Linked releaseLin’s follow-up links ARC Prize’s Argon announcement, Astra results page, scoring methodology and vendor documentation about preserving reasoning context.
Source follow-up retained with the root; edition desk should verify causal wording against primary evidence before reuse.
Posted .
Alan Lin discusses the previously published ARC-AGI-3 standard/provider-adapter results and argues that retaining reasoning between steps matters when choosing a harness.
Fresh public commentary on existing results. Treat as Lin’s interpretation, not a new evaluation or proof that reasoning retention alone caused the difference.
Posted .
Replying to a question about agents relearning project context, Puget points to Graphify as one of several tools.
General agent-workflow context from an ARC contender; no performance claim.
Posted .
Linked releaseReplying to a suggestion to inspect checks rather than agent-written code, Puget says that approach cannot detect overengineering.
Attributed workflow caution; no ARC scoring claim.
Posted .
Itsu posts the Japanese word ゴイサギ (black-crowned night heron), #kaggle, and an ARC-AGI-2 competition link.
Cryptic bird reference; no explicit score, rank or medal in the actual post. Do not infer a standings claim from the pun.
Posted .
Cole thanks Dries Smit for a supportive reply about his addition to the ARC-2 team.
Small reaction to an already reported team announcement; does not establish a new roster change.
Posted .
Klambauer announces KronLSTM and links the new Kronecker Delta Rule paper. The authors propose a targeted association update for recurrent linear attention, report competitive long-context performance, and link open-source code and models. No ARC score is reported.
Research lead discovered in fses91’s separate Reposts feed. The primary arXiv abstract was checked independently; architecture claims remain attributed, and no Kaggle roster identity is inferred.
Posted .
Linked releaseTufa Labs says Dominique Garmier and Andrin Rehmann will attend NVIDIA GTC Berlin this week and invites attendees to arrange a meeting through the lab.
New public team activity from the leaders of both ARC boards; no score or research-method claim.
Posted .
Knoop challenges Rohit Krishnan’s prediction that judgment and taste will remain hard to generalize. He says asking interesting open questions has not been formalized, but believes it could be made tractable if formalized. This is an attributed research opinion, not a benchmark result.
Organizer commentary relevant to open-ended discovery and evaluating interesting questions. Actual reply and expanded parent read in signed-in Chrome; no subscriber restriction on either.
Posted .
Geremie Yeo says he will attend the PyTorch conference on October 20 and 21.
Contender community colour; no standings change claimed.
Posted .
The author contrasts the old process of navigating Google results and ads with present AI-assisted information seeking, and wonders what the next few years will bring.
Captured because newly visible in Matt Mazur’s Reposts inventory. Original public post read. Repost action time is unknown; this is attributed general AI commentary, not an ARC result.
Posted .
Translation of original Japanese: I improved my best score, though only a little. Following ARC3, I am aiming for the gold-medal band in ARC2 too. The Kaggle share text says rank 40.
Original Japanese read on actual public post. Direct X → nebosukeai.com → Kaggle isakatsuyoshi link verified, plus reciprocal Kaggle → X link. Public ARC2 board independently shows rank 40, 44.44, team 復活の混テキスト; gold is an ambition, not an awarded medal.
Posted .
Mazur offers his arXiv paper-content collection and analysis APIs to potential collaborators, replying to Burny.
Public research collaboration lead; no new ARC result or competition claim.
Posted .
Mazur asks Dave Martin when he expects to launch Crafd, in reply to a side-business progress update.
Organizer capture for completeness; unrelated to ARC and not a news lead.
Posted .
Malysheva says the Mostik offsite prompted a stricter documentation habit: anything that came up twice should go into a document, so the team retains what it learned together.
Attributed team-working colour; no benchmark result or score claim. Full actual public post read in signed-in Chrome.
Posted .
Cole publicly confirms joining Mark Barney’s ARC-AGI-2 team, says he is excited, and says they will work hard through the last two weeks because the competition is strong. His quote contains Barney’s public ARC2 team announcement.
New first-person team confirmation. Signed-in Chrome original post read; public post has no subscriber restriction. Exact public roster also shows jcole75 with markbarney, sonphamorg and scizical in ARC Mutants; this does not establish a joining date or score contribution.
Posted .
Mark Barney publicly welcomes Jack Cole to his ARC-AGI-2 team, using a Professor X metaphor, and states an ambition to reach gold. The post has ARC Mutants artwork. The announcement explicitly concerns ARC2; it does not add Cole to ARC3 Team Turducken.
Fresh public first-person confirmation supporting the commissioned ARC Mutants dispatch. Gold is an ambition, not an awarded medal.
Posted .
Replying to Dmitrii Kovanikov’s comment that agents wait for instructions, Mazur says recurring routines can give Claude or Codex ongoing work. He says his experiments over recent months continue to surprise him and quotes his earlier Preceden example.
General AI workflow commentary from an organizer; attributed experience, with no new ARC benchmark or Kaggle result. Actual reply and parent conversation read in signed-in Chrome.
Posted .
In a SKY-001 follow-up, Pasupalak says his team tested CAD and BIM tasks against GPT-6 Astra, Claude Opus 5.5 and other frontier models and claims its model outperformed them. He argues autoregressive LLMs are ill-suited to visual reasoning and long-horizon tasks.
Context for the release claim; performance and architectural conclusions are attributed and unverified. Posted after scan start and captured opportunistically.
Posted .
Sam Pasupalak announces SKY-001, which he calls the first Engineering World Model. He claims higher IoU on CAD and BIM tasks than tested frontier models, with 51–125 times faster and 150–210 times cheaper operation. Says it reflects 18 months of work by a team from Maluuba. Observed in Chollet’s repost inventory.
Adjacent research lead amplified by an organizer; author claims only, with no independently inspected evaluation or ARC score. Repost action time is unknown.
Posted .
In his ChatGPT Then and Now launch conversation, Mazur shares a humorous generated Cinderella retelling in a Donald Trump style and a link to that comparison.
Attributed humour and product colour; no ARC contest claim. Actual reply and parent conversation read.
Posted .
Replying to a reader reacting to the grain-of-sand story in ChatGPT Then and Now, Mazur supplies the direct comparison link.
Follow-up to the already archived comparison-tool launch; no new benchmark result. Actual reply and parent conversation read.
Posted .
ARC Prize links its leaderboard, public reproduction repository, testing policy and full Beam 1.0 results.
Primary follow-up with reproducibility and policy links; group with the Beam announcement.
Posted .
Linked releaseARC Prize says Beam costs use Baseten pricing of $1.30 input and $4.30 output per million tokens. ARC-AGI-3 results are pending.
Keeps the cost basis and absence of ARC-AGI-3 results explicit.
Posted .
Linked releaseARC Prize reports verified Reflection Beam 1.0 scores: ARC-AGI-2 23.5% at $0.61/task and ARC-AGI-1 74.0% at $0.31/task. Best scores use xhigh reasoning; high reaches 20.0% on ARC-AGI-2 at about 43% lower cost.
New official model evaluation. Linked result page independently read in Chrome: semi-private best scores match. This is benchmark evaluation, not a Kaggle team score.
Posted .
Linked releaseReplying to CaseySoftware about autonomous software development, Mazur says he orchestrates his work through the Claude Code desktop app on his Max subscription without extra charges. He asks about the reported $90 cost, mistakes and surprises.
Organizer commentary on agent workflows; general community colour, not an ARC result or score claim.
Posted .
Replying to Lane Henslee in the ChatGPT Then and Now launch thread, Mazur compares the HTML-for-a-dog answers: both use a park analogy, while Astra adds formatting and emoji. The linked public example was read and supports that narrow comparison.
New organizer thread follow-up; a qualitative example, not an ARC result.
Posted .
Linked releaseReplying to a suggestion to inspect only some parts of agent code, Puget asks how one can know what the agent changed without checking.
Follow-up on human review of experimental code.
Posted .
In the workflow thread, Puget says his priorities are human-readable code and performance when responding to a proposed codebase skill.
Follow-up context for his agent workflow commentary.
Posted .
Mazur recalls early ChatGPT hallucinations being indexed and reused in Google quick answers, links his October 2023 Kenya example, and offers a present-day Astra comparison.
Follow-up context about hallucinations; this run did not independently audit the historical search behavior.
Posted .
Linked releaseThread follow-up links Karpathy’s build-GPT tutorial and says LearnGPT appears near the one-minute mark.
Context for the release; historical tutorial claim attributed to Mazur.
Posted .
Linked releaseMazur launched ChatGPT Then and Now, pairing 492 early LearnGPT responses with new GPT-6 Astra answers. The linked page identifies the originals as December 2022–March 2023, new runs as October 19, 2026, one message with no system prompt and default reasoning effort, and says Claude Opus 5.5 wrote the paired analyses. A data download is offered; it was not downloaded.
Public organizer product/research lead; a prompt comparison collection, with no ARC score or competition result claimed.
Posted .
Linked releaseChollet observes that widely shared expectations can produce consequences even if the anticipated event never occurs.
Attributed general commentary; no ARC result or explicit prediction.
Posted .
Cole congratulates Mark Barney on his ARC-AGI-3 top-three post and jokes that the top two might be nervous. The quoted post and actual public quote were read.
Friendly contender reaction to an already reported standings event; no new board result.
Posted .
Tufa says its two NVIDIA GB300 NVL72 racks are live in Stockholm; credits DELTA for installation and T.Loop for hosting and links a deployment video. Actual public post read.
Contender infrastructure announcement; no evidence here ties this equipment to a specific score or competition run.
Posted .
Puget says AI agents now run all his training experiments, while he still reads their harness code. He says agents tend to overengineer, retain unnecessary compatibility code and require human direction. Actual public post expanded and read.
First-person account of a competitor’s research workflow; no new ARC score claimed.
Posted .
Anders Rippe says his first personal harness submission reached rank 626 on ARC-AGI-3. He jokes about needing more RAM. The shared link resolves to the ARC-AGI-3 contest; no exact Kaggle handle or team identity has been verified.
New participant-led harness progress claim for community coverage; self-reported rank, not an independently matched board result.
Posted .
Linked releaseMazur thanks @lordofscripts for sharing a Panama presidents timeline made with his Preceden AI Timeline Maker.
Routine product acknowledgment, captured before editorial selection; no new ARC result or research claim. Actual post and parent context read in signed-in Chrome; full reply independently visible while logged out.
Posted .
Replying to a question about his Preceden automation, Mazur shares an answer attributed to Claude describing scheduled desktop tasks with separate playbooks, a shared rulebook, disk-based state, bounded changes, tests and review gates, separate deployments, and a weekly retrospective with human decisions.
Useful organizer commentary on agent orchestration, captured before editorial selection. Describes his own software workflow; makes no new ARC benchmark or competition result claim. Full text visible on the actual post while logged out.
Posted .
Malysheva links two studies arguing that fewer tokens do not reliably mean lower agent cost or latency, and connects this to Mostik’s latent-bridge approach. Both paper abstracts were read. The first was revised in August and the second submitted in September; this is a new post about existing research, not new ARC results.
Useful inference-budget research angle. Primary abstracts support the narrow compression findings; her Mostik connection is her interpretation and promotion, not an ARC score claim.
Posted .
Linked releaseReplying to Psyho’s PuzzleScript benchmark announcement, Barry says Epoch might be able to support API costs and that he has sent a DM. This is a tentative public statement; no funding agreement or collaboration is established.
New response to the archive’s independent benchmark lead; recheck any later public confirmation before describing support as secured.
Posted .
Kamradt publicly replies that he sent a DM to Psyho after Psyho announced a nearly finished private PuzzleScript-based reasoning benchmark.
Organizer interest in an independent game benchmark; the message contents and any collaboration are unknown.
Posted .
Asked about online solutions and prior model knowledge, Psyho says the selected games are niche and that reviewing logs for all selected games revealed no suspicious behavior.
Author-reported contamination check; does not independently establish absence of training exposure.
Posted .
Psyho reports full-run costs of $700 for Sol 6.1 and $60 for Luna 6, and estimates about $1,000 for Sonnet 5.5 and $4,000 for Astra 6. Says testing uses xhigh and may reduce generous action limits to lower cost.
Separates reported completed costs from forecasts; no logs, scores or reproducible release verified.
Posted .
Psyho jokes that finishing the benchmark is procrastination from preparing for an interview.
Attributed colour accompanying the announcement.
Posted .
Psyho says a nearly finished PuzzleScript benchmark selects 40 games from over 300, uses Codex/Claude Code harnesses with code generation prohibited, and keeps the game set private. The author says generous action limits already allow some saturation while models remain inefficient.
Potential benchmark-development story; author claims, not an official ARC release or independently verified score.
Posted .
Mazur replies “Well said!” to Zanthous’s objection that knowing what a benchmark is like does not itself compromise it.
Adds an organizer reaction to the existing debate about synthetic practice environments; supplies no new measurement.
Posted .
Kamradt jokes about a VC asking for a business moat, quoting Musk’s claim about superintelligence, robots, solar power and mass to space.
Organizer colour; attributed joke and quoted prediction, not an ARC result.
Posted .
Zanthous asks whether merely knowing what a benchmark is like is enough to call it compromised, and calls that standard excessive.
The direct parent of Mazur’s approving reply.
Posted .
Zanthous argues that private novel games remain meaningful even if labs create practice environments, because skill transfer is not guaranteed.
Context for Mazur’s new approving reply; an attributed argument, not proof about any lab’s training.
Posted .
Graham predicts persistent agent attacks while quoting dax’s report of attacks over the past 48 hours and a need for defenders running continuously.
Newly observed in Mazur’s repost inventory. General AI/security commentary; repost time unknown and attack claim not independently verified.
Posted .
Marc Shade self-reports rank 766 in an ARC-AGI-3 Kaggle share, joking that it is a small price for a social life.
New public community colour. The competition link is verified, but Kaggle handle, team ID and current rank are unverified.
Posted .
Linked releaseIn reply to Itsu, NT says their side is in the same situation and jokes that there is heavy traffic between 60 and 70.
New thread context about difficulty improving; no Kaggle identity or team membership established.
Posted .
Linked releaseReplying after NT congratulates them on passing 70, Itsu says they do not yet know how to improve beyond 70 and will keep trying.
Attributed contender reaction, with thread context; no method or new official result established.
Posted .
Linked releaseNT congratulates Itsu on exceeding 70 points in the ARC-AGI-2 thread.
Context for Itsu’s already archived response; not an independently checked score.
Posted .
Linked releaseItsu says they strongly expect to fall from rank eight on the private leaderboard, and perhaps on the public board too. This is their expectation, not a private result. The post links ARC-AGI-2.
New first-person reaction from a top-ten contender; distinguish the rank mentioned at posting from later board observations.
Posted .
Linked releaseGeremie Yeo reports a 12-minute-59-second wait to provision four L4 GPUs on Modal.
A dated personal report following his October 17 Kaggle-queue commentary. It does not establish a platform-wide wait or attribute a particular ARC score to this run.
Posted .
Mazur replies “Good question!” to Raikaru asking what benchmark could not possibly be trained for, in the ongoing benchmark-training discussion. No new testing claim.
Organizer follow-up archived for completeness; colour only, no substantive new result.
Posted .
Mazur replies “same :)” to Krzysztof Cichy joking that the Preceden experiment makes him want to give AI more control.
Organizer follow-up archived for completeness; colour only, no substantive new result.
Posted .
Mazur says that, based on conversations with some lab staff and to the best of his knowledge, none have built synthetic RL environments specifically to raise ARC scores. He explains that ARC calls datasets semi-private because evaluation data leaves the organization through third-party model APIs, despite zero data retention.
Follow-up to the evaluation-integrity exchange. Attribute to Mazur; this is not an independent audit of lab training data. The critic acknowledges having no evidence of that practice.
Posted .
Mazur asks whether there is evidence that major labs create synthetic ARC-like reinforcement-learning environments specifically to improve ARC benchmark scores.
Follow-up to the evaluation-integrity exchange. Attribute to Mazur; this is not an independent audit of lab training data. The critic acknowledges having no evidence of that practice.
Posted .
Chollet disputes a claim that AI is improving software-company profits. He says software-engineer employment is flat at SaaS firms and revenue follows trends predating AI, citing Adobe, Workday, Zoom and Salesforce. Four attached charts were visible; their data were not independently verified.
Attributed follow-up to his earlier software-automation commentary; opinion and empirical claims, not a benchmark result.
Posted .
Knoop recalls 1998 Zelda Online / Graal as an unmatched experience, says he was a GC, and criticizes its management.
Organizer colour captured before filtering; no contest or research development.
Posted .
Responding to an allegation that benchmarks test on training data, Mazur says ARC is not compromised, its semi-private evaluation sets have not leaked and cannot be trained on, and provider API testing requires zero data retention.
Useful organizer statement on evaluation integrity; attribute to Mazur rather than presenting it as an independent audit.
Posted .
Mazur praises a pathfinding visualization shared by Binx, replying “Cool viz”. The parent describes an older random pathfinding comparison.
Organizer community colour; no competition score, release or research result.
Posted .
Mazur agrees with Mark Barney’s remark about the interaction of language and mathematics, replying “Same” in the Preceden discussion.
Organizer community colour; no competition score, release or research result.
Posted .
Mazur asks Billy to explain algorithmic disintermediation, in a discussion quoting his autonomous SaaS coding experiment.
Organizer community colour; no competition score, release or research result.
Posted .
Mazur replies with a hundred-points emoji to Charles Smith’s praise of Claude finding a broken funnel and knowing where to stop. A brief agreement, with no new technical detail or ARC result.
Captured before editorial selection; colour only.
Posted .
Mazur enthusiastically agrees with Lane Henslee’s praise of the existing-subscription setup and the old thumbnail bug the agent reportedly found.
Organizer commentary captured before editorial selection; attributed claims, not an ARC benchmark result.
Posted .
Chollet argues that coding automation shifts bottlenecks elsewhere in a system. He claims engineering demand and workloads remain high despite agents writing most code, while companies often see little resulting profit improvement. His commentary gives no supporting study or ARC score.
Organizer commentary captured before editorial selection; attributed claims, not an ARC benchmark result.
Posted .
Mazur links back to his October 9 account of Claude Code finding an old SaaS bug, submitting and deploying a fix, and measuring affected production records. This is a new reply referring to an earlier claimed result, not a new fix.
Organizer commentary captured before editorial selection; attributed claims, not an ARC benchmark result.
Posted .
Linked releaseMazur tells Mark Barney that the Preceden date-localization bug is fixed. He says a French language selection combined with an English topic produced English content with French dates, which the homepage cache then served to everyone. He says the language and caching logic needs simplifying.
Organizer commentary captured before editorial selection; attributed claims, not an ARC benchmark result.
Posted .
Mazur responds with a cheers emoji after Pavel Flit acknowledges that AI-written replies may reduce genuine engagement.
Organizer commentary captured before editorial selection; attributed claims, not an ARC benchmark result.
Posted .
Mazur argues that visibly AI-generated replies discourage engagement, even when the underlying ideas belong to the writer.
Organizer follow-up captured before editorial selection; no ARC score or standings claim.
Posted .
Mazur asks a commenter to reply in their own words rather than use an LLM to write replies.
Organizer follow-up captured before editorial selection; no ARC score or standings claim.
Posted .
Mazur thanks Mark Barney for suggesting an independent daily QA agent acting as a new user and says he will brainstorm the idea.
Organizer follow-up captured before editorial selection; no ARC score or standings claim.
Posted .
Mazur responds with agreement and a grimacing emoji to a comparison of the agent’s reported record with human product managers.
Organizer follow-up captured before editorial selection; no ARC score or standings claim.
Posted .
Mazur says the system uses his existing $200/month Claude Code subscription, adds no extra bill, and barely affects weekly usage. He clarifies that routines start hourly and often finish within minutes, rather than running continuously.
Organizer follow-up captured before editorial selection; no ARC score or standings claim.
Posted .
Mazur posts an answer attributed to Claude: signup-funnel data identified a weak cohort, a desktop/mobile browser walkthrough revealed that generated timelines exceeded the free event limit, and code inspection explained the disabled Add Event button. The reply says plan logic required Matt’s approval and a fix is scheduled for Monday.
Useful detail behind the Preceden autonomous-agent experiment. Author-published account, not independently verified performance or a shipped fix.
Posted .
Knoop jokes by linking the Wikipedia entry for cute aggression in response to a proposal to align AI by making it regard humans as cute.
Organizer follow-up captured before editorial selection; no ARC score or standings claim.
Posted .
Linked releaseResponding to a warning that conversion gains can mask a decline in willingness to pay, Mazur says this is not too hard to guard against; he gives no method in this reply.
Organizer commentary captured before editorial selection; no new ARC score or independently verified product outcome.
Posted .
Linked releaseAsked whether Claude has shipped anything he regretted, Mazur says it has not and praises its competence. This is his personal experience claim.
Organizer commentary captured before editorial selection; no new ARC score or independently verified product outcome.
Posted .
Linked releaseKamradt replies to Michael Tan's North Beach/San Francisco post with a link to PixieCitie, his live city-view app.
Organizer commentary captured before editorial selection; no new ARC score or independently verified product outcome.
Posted .
Linked releaseMazur thanks @fireandstart for sharing his autonomous Preceden development experiment.
Organizer commentary captured before editorial selection; no new ARC score or independently verified product outcome.
Posted .
Asked how much of the agent-generated backlog he ships, Mazur says he follows Claude's recommendations about 99% of the time.
Organizer commentary captured before editorial selection; no new ARC score or independently verified product outcome.
Posted .
Linked releaseReplying to a comment favoring human approval for pricing, limits and authentication, Mazur says funnel autonomy could be fairly low risk if the agent measures changes and iterates. He frames lower revenue over days or weeks as a downside and lasting conversion and revenue gains as a possible upside.
Attributed view on autonomy and outcome measurement; hypothetical benefits, not a demonstrated revenue gain.
Posted .
Mazur links his September 21 account of experimenting with Claude Code autonomously shipping changes to Preceden as background to the new funnel thread.
Context link for the autonomous coding experiment; retain with the root thread.
Posted .
Mazur says his autonomous Claude setup found a signup funnel problem in Preceden without being directed to it: AI-generated timelines exceed the free plan limit, leaving new users unable to add events. He says it investigated production data and code, proposed fixes to limits, signup thumbnails, upgrade copy and tracking, and defined a 28-day success check. Pricing, plan limits and authentication changes required his approval.
Organizer perspective on agent autonomy and measuring product outcomes; an attributed product anecdote, not an ARC score or independently verified business result.
Posted .
Linked releaseIn Japanese, @menphim says: “I am in the ARC-AGI-3 silver band for now, but I do not understand anything. Is it possible fewer than 50 people understand this competition?” The final question is a rhetorical reaction, not a measured headcount.
Participant colour about the contest’s difficulty. The silver-band position is self-reported here; no new roster or standings claim established.
Posted .
Geremie Yeo complains about Kaggle inference queues and competition demand for four L4 GPUs. His post mentions both ARC contests and Gemma 4 Developer Agent.
A dated participant reaction to compute contention; his explanation and cloud-price claim are not independently established capacity measurements.
Posted .
Mazur replies to a comparison of modern tutorial-heavy games and 1990s games by revisiting his September Contra III AI-modding experiment.
Gaming and AI colour; this is a retrospective reference, not a new model or benchmark result.
Posted .
Mazur expresses nostalgia in response to a post showing a familiar game map.
Personal gaming colour; no competition claim.
Posted .
Mazur joins a nostalgic exchange about an old poker software client.
Personal banter; no competition or research claim.
Posted .
Murphy says his ARC talk next week in Boston will cover his Model Discovery Agent and an interactive scientific benchmark inspired by ARC-AGI-3. Agents would run a few experiments in a partially observed stochastic simulator, fit a causal world model, then be evaluated on predictions for novel interventions rather than on the model description itself.
A new primary-source lead for Research Summit coverage and the next generation of interactive benchmarks. A proposed benchmark and upcoming talk, not a released benchmark, new score, or Kaggle result.
Posted .
In Japanese, masspeaks thanks NT (@Mitubant) and jokes that he looks forward to continuing their matches in both ARC-2 and Street Fighter 6. The parent conversation celebrates a Kaggriculture medal; that result is not an ARC medal.
Friendly contender colour only. No new ARC score, medal, identity or team-membership claim.
Posted .
Mazur encourages Dave Martin in a reply to Martin’s post about coming to enjoy running.
Organizer colour only; no ARC result or research claim.
Posted .
Replying to Binx’s “more from the primordial soup” video, Mazur calls it cool, asks how it was built and tags Fraser Price.
Routine organizer community colour retained before editorial selection; no new competition result or research release.
Posted .
Replying to Hrishi about the Emergent Mind homepage card-title font, Mazur says Opus chose it and identifies it as Menlo.
Routine organizer community colour retained before editorial selection; no new competition result or research release.
Posted .
Linked releaseReplying to Finn Ramos in the Emergent Mind homepage conversation, Mazur invites him to explore and report issues or suggestions.
Organizer community follow-up retained under capture-before-filtering; no new research or competition result.
Posted .
Linked releaseMazur encourages Fraser Price to continue his artificial-life experiments and predicts they will attract a larger audience. Encouragement, not a validated research result.
Organizer community colour; no new contest standing or benchmark result.
Posted .
In the Emergent Mind relaunch conversation, Mazur says Fraser Price’s similar experimental approach partly inspired him.
Organizer community colour; no new contest standing or benchmark result.
Posted .
Daviswh points to the different Argon ARC-AGI-3 results under the provider adapter and standard harnesses and says harness choice should accompany leaderboard claims. Reaction to ARC Prize’s existing announcement, not a new evaluation.
Optional community colour for the harness story; no new result, competitor identity or recipe established.
Posted .
Linked releaseIn reply to a reader praising the homepage, Mazur says it now matches what he always wanted the site to be.
Organizer capture before editorial selection. Product and community context, not an ARC score or competition result. Each item was read at its actual permalink with no subscriber restriction.
Posted .
Linked releaseMazur thanks a reader and says he is adding the reader’s suggested experiment to his list; this is a conversational response, not an announced release.
Organizer capture before editorial selection. Product and community context, not an ARC score or competition result. Each item was read at its actual permalink with no subscriber restriction.
Posted .
Linked releaseResponding to Dave Martin encouraging him to keep making things, Mazur says he always has and always will.
Organizer capture before editorial selection. Product and community context, not an ARC score or competition result. Each item was read at its actual permalink with no subscriber restriction.
Posted .
Linked releaseMazur confirms that keeping the assistant as the first featured tool was intended to let existing users find it quickly.
Organizer capture before editorial selection. Product and community context, not an ARC score or competition result. Each item was read at its actual permalink with no subscriber restriction.
Posted .
Linked releaseMazur thanks a reader for congratulations on the Emergent Mind relaunch.
Organizer capture before editorial selection. Product and community context, not an ARC score or competition result. Each item was read at its actual permalink with no subscriber restriction.
Posted .
Linked releaseMazur thanks a reader who compared discovering his site to browsing early-2000s experimental websites.
Organizer capture before editorial selection. Product and community context, not an ARC score or competition result. Each item was read at its actual permalink with no subscriber restriction.
Posted .
Linked releaseMazur clarifies that the old homepage was the arXiv AI research assistant; it remains the first featured tool and is available at emergentmind.com/assistant.
Organizer capture before editorial selection. Product and community context, not an ARC score or competition result. Each item was read at its actual permalink with no subscriber restriction.
Posted .
Linked releaseMazur posts an image/GIF reaction to a reader’s brief praise of the Emergent Mind launch; no text or measurable result is stated.
Organizer capture before editorial selection. Product and community context, not an ARC score or competition result. Each item was read at its actual permalink with no subscriber restriction.
Posted .
Linked releaseMazur thanks Michael Byrne for calling the site an amazing portfolio and says it is coming along.
Organizer capture before editorial selection. Product and community context, not an ARC score or competition result. Each item was read at its actual permalink with no subscriber restriction.
Posted .
Linked releaseMazur announces a new Emergent Mind homepage and a shift to an experimental product lab, with the arXiv research assistant retained as one of its tools. He says most tools will be small and free, with some larger paid tools. The actual homepage independently displays AI Product Lab by Matt Mazur and links the assistant, open problems and experimental projects.
Organizer capture before editorial selection. Product and community context, not an ARC score or competition result. Each item was read at its actual permalink with no subscriber restriction.
Posted .
Linked releaseItsu announces a move to rank 14 in ARC-AGI-2, with a joke comparing the climb to the electricity bill. The original post links the ARC-AGI-2 competition. Chrome subsequently showed Muggle at rank 14 with nisshokuitsuki as its listed account; the collector snapshot at 09:49:46 UTC maps that account to team 16297545.
Fresh contender statement for the Early edition: rank 14 is within the observed provisional public gold band. Do not imply the score was achieved at the post time or a final medal has been won.
Posted .
Linked releaseJack Cole congratulated Mark Barney in a reply to the Digest announcement that Son & Mark & Ronen had reached third on ARC-AGI-3.
Friendly competitor reaction; not evidence of a new score, team affiliation or method. Actual permalink and parent read in signed-in Chrome; no subscriber restriction shown.
Posted .
Translated from Japanese: fshindo says trying Nemotron 3.5 made him appreciate how capable the Qwen 3.8 family is.
Personal model-use opinion; no ARC score, harness, training or competition-method claim. Actual post read; no subscriber restriction shown.
Posted .
In reply to Mark Cuban’s claim that AI is book smart rather than street smart, Mazur asks how that distinction would be measured.
General commentary on evaluating AI; no new ARC result or standings claim.
Posted .
In the Argon announcement thread, Greg asks a commenter criticizing graphic quality for specifics and says the team is happy to improve it.
Captured before editorial filtering; minor community feedback, not a revision to the result.
Posted .
Linked releaseIn the same Argon announcement thread, Mazur thanks the graphic-quality critic and says the team will work on it.
Captured before editorial filtering; community follow-up, not a score correction.
Posted .
Linked releaseReacting to Meta’s Muse Spark Open announcement, fshindo asks in Japanese whether the UD-IQ2_XXS quantization might just fit on two Sparks.
An attributed hardware feasibility question, not a tested fit, benchmark, or ARC competition result.
Posted .
Linked releaseNick adds that agents have automated circuit-study work that previously took months for the Curve Detectors / Curve Circuits experiments on ten neurons.
Attributed claim about research automation, not a measured ARC result.
Posted .
Linked releaseResponding to Knoop, Nick says such an evaluation seems feasible and points to controlled activations, ablations, large token collections and other circuit-study tools; he promises more details.
Substantive response to the organizer’s evaluation proposal. Feasibility is the author’s view.
Posted .
Linked releaseReplying to Fraser Price’s simulated multicellular-life video, Mazur praises the research but says the difficult next step is advancing beyond such demonstrations, perhaps by evolving rudimentary intelligence.
Organizer commentary on open-ended research. The parent simulation claim was read but not independently validated.
Posted .
Linked releaseReplying to Nick’s Neural Networks Are Already Clean announcement, Knoop asks whether neural-network cleanliness can be formalized as an evaluation and whether the proposed falsifiable-explanation experiment would suffice.
Concrete evaluation question, with two substantive replies from the original author. No ARC score or validated new evaluation announced.
Posted .
Linked releaseKnoop argues experiment ideation is the current bottleneck in recursive self-improvement. He claims coding and execution are largely automated, distinguishes reproducing known results from asking novel questions, and says domain-specific improvement recipes still need human-AI collaboration to spread across domains without further generalization breakthroughs.
Useful organizer perspective on open-ended evaluation and ARC-AGI. His automation percentages and 12-month scenario are attributed claims, not independently verified results.
Posted .
Chollet says Argon has a similar ARC-AGI-3 performance profile to Astra and calls Google back at the frontier, quoting ARC Prize’s result.
Attributed organizer reaction to the Verified release; not a separate evaluation.
Posted .
Linked releaseItsu reports moving to rank 30 on Kaggle. The original post links ARC Prize 2026 - ARC-AGI-2.
Dated competitor statement. Later collector snapshot at 20:48:28Z lists Odd Dice, team 16297545, roster nisshokuitsuki, at rank 32; the post is not a current-rank claim.
Posted .
Linked releaseKnoop calls Argon the second model family effectively to max out ARC-AGI-3 with a provider adapter and notes a similar backwards bend in action efficiency.
Attributed reaction; preserve the provider-adapter qualification and avoid equating it with Kaggle performance.
Posted .
Linked releaseARC Prize links the Verified leaderboard, reproducible public benchmark code, testing policy and full Gemini 4 Argon results.
Primary evidence links for edition follow-up; no new Kaggle ranking.
Posted .
Linked releaseARC Prize says API issues prevented complete standard-harness testing: low reasoning completed 42/55 games for 49.8%; medium completed 46/55 for 46.3%; high completed 19/55 and no score was published. Reported scores retain partial credit across all 55 games. Provider adapter preserves opaque reasoning and permits automatic compaction; standard harness carries notes between turns. Costs use introductory $2 input/$10 output per million tokens.
Material limit on harness comparisons. Do not present incomplete standard-harness runs as a clean capability ceiling.
Posted .
Linked releaseARC Prize places Argon behind Astra’s 99.9% ARC-AGI-3 result. It reports 62% fewer actions than the human baseline on the 25 public games, matching or beating three speedrun records and coming within 10% on six others.
Keep the public-demo action comparison separate from the semi-private Verified score.
Posted .
Linked releaseARC Prize announces Gemini 4 Argon Verified results: ARC-AGI-3 97.5% at $4.8K with the provider adapter and 49.8% at $5.8K with the standard harness; ARC-AGI-2 91.3% at $0.32/task and ARC-AGI-1 98.5% at $0.16/task. It calls Google DeepMind the second frontier lab above 95% on ARC-AGI-3.
Consequential organizer result. Separate from Kaggle standings. Standard-harness API incompletion is essential context; primary results introduction says 91.2% on v2 while its table and announcement say 91.3%.
Posted .
Linked releaseJoshi introduces SHELDON, a reasoning harness that separately tracks hypotheses and observations. His thread claims ARC-AGI-2 improvement over its baselines. The linked Stanford project table reports 59.17 for SHELDON versus 47.50 for EvoX, using GPT-5.6-terra high reasoning and GPT-5.6-luna for delegated implementation.
Fresh primary research lead. These are author-reported experimental results, not Kaggle standings or an ARC Prize Verified evaluation; confirm split, budget and comparability before publication.
Posted .
Linked releaseMalysheva discusses RaReCache, which maps a smaller model’s prompt KV cache to a larger model and selectively recomputes critical token positions. The primary abstract describes rank disagreement selection, retained target accuracy and reduced prefill cost.
Existing archived lead refreshed against the October 8 paper abstract; no new ARC score or independently reproduced speed claim.
Posted .
Linked releaseChollet recommends Gakku, describing a voice tutoring app that coaches school-age students through short essay-writing missions. He argues conversational AI could make individual tutoring widely accessible.
Organizer commentary captured before editorial selection; an education recommendation, not an ARC benchmark result.
Posted .
Linked releaseChollet argues that breakthroughs come from sustained interest in the process of solving hard problems, rather than merely wanting the outcome.
Organizer general commentary; attributed opinion, not a measured result.
Posted .
Translation of Japanese original: An ARC2 merge invitation involving the gold range has come in, so I am torn, but I want to keep trying solo until the last possible moment. The post does not name the inviting team or announce an accepted merger.
Contender motivation and possible future team change; do not report a merger or a verified gold position. Exact current team mapping remains unresolved.
Posted .
Knoop says taste is about asking interesting questions. This is his general commentary, not a competition result.
Organizer capture before editorial selection; minor colour only, with no standings or benchmark implication. Actual post opened in signed-in Chrome; full text visible without a Subscribers label or access restriction.
Posted .
Replying to Vlad Garbuz’s question about the provider checkpoint used for the DeepSeek evaluation, Mazur says Baseten serves the original DeepSeek V4.1 Flash checkpoint, which natively uses fp8 with mxfp4 experts, without additional quantization. Original reply and parent chain read in regular public-post UI, with no Subscribers restriction. This is Mazur’s testing clarification, not an independently reproduced provider comparison.
New organizer detail for interpreting the October 13 verified DeepSeek evaluation. Keep the native mixed precision distinct from additional provider quantization; no new Kaggle score or ruling.
Posted .
Plain-English translation: “Opus 5.5 suddenly started speaking only English and it’s irritating. This suddenly happens in every session.” The public original reports the author’s experience. Nearby replies speculate about degradation; no provider change or quantization finding is established.
Japanese competitor background about model use. This is an anecdote, not an independently verified product regression or contest recipe.
Posted .
Plain-English translation of the original Japanese: “ARC-AGI-3: I’m not seeing any improvement locally either, and I have no idea what to do.” This is the author’s own progress frustration; neighboring cross-validation/day-six updates concern other work and are not assigned to ARC.
A current Japanese competitor voice for desk context, not a score claim or verified failure of a particular method.
Posted .
ARC Prize supplies its leaderboard, reproduction repository, testing policy and full Haiku 5.5 results as supporting links for the evaluation thread.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Linked releaseARC Prize says it omitted Haiku 5.5 max reasoning on ARC-AGI-2 because only 74 of 120 tasks completed, with many incomplete attempts exhausting their output budget before producing parseable answers.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Linked releaseARC Prize reports verified Haiku 5.5 scores: ARC-AGI-3 2.8%/$437 with its provider adapter and 0.5%/$388 with its standard harness; ARC-AGI-2 61.3% at $0.085 per task; ARC-AGI-1 89.0% at $0.051 per task. These are model evaluations, separate from Kaggle team standings.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Linked releaseReplying to Daniel Tenner, Mazur suggests people can develop taste and vision rather than treating them as fixed traits.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Mazur asks Daniel Tenner how people can acquire taste and vision, in a discussion of AI-assisted work.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Chollet argues that autonomous robotics needs causality and discrete reasoning beyond autoregressive sampling, and praises Biwei Huang’s Aether AI work. His quote refers to Huang’s October 12 CRIS-0 announcement; it is not an October 14 model release or a Kaggle result.
New public founder research lead. Original and visible conversation read; public body independently readable without sign-in. Linked CRIS-0 technical project page read: explicit task state, verification, retry/replan; reported robotics work, no disclosed ARC recipe.
Posted .
Linked releaseIn a direct reply to Biwei Huang’s October 12 announcement, Chollet congratulates the CRIS-0 work.
Mandatory capture of founder reply; public original/date and parent context checked in logged-out Chrome. Group with his later causality quote.
Posted .
Linked releaseChollet uses a pointed string-theory analogy to argue that intelligence requires grounding and repeated contact with reality, which he sees as the practical bottleneck in solving important problems. This is his commentary, not an established scientific result.
Mandatory founder capture; complete original body and author date readable in logged-out Chrome. No new ARC standings claim.
Posted .
Japanese translation: Odd Dice’s flick/slide input judgement was so strange that I deliberately chose it as my Kaggle team name. A light explanation of the name, not a disclosure of an ARC solution or model method.
Own team-name statement complements direct Kaggle-to-X identity evidence. Original body/date readable without sign-in; preserve the joke as color.
Posted .
Japanese translation: Even with a 2.5-point improvement, I have actually dropped from rank 31. The attached native Kaggle share says Moved up to rank 35 for ARC-AGI-2. This is the author’s posting-time comparison, distinct from the saved rolling board event and later rank.
Own standings voice. Kaggle nisshokuitsuki directly links @iiiiitsu_ne; current ARC-2 Odd Dice roster contains that handle. Original body/date independently readable without sign-in.
Posted .
Chollet argues that the ideal scientific outcome of AI is to enable millions of human scientists to pursue their curiosity more ambitiously, rather than concentrating expensive proprietary systems in a few labs tackling prestige problems.
A fresh public founder statement at 11:43 pm Eastern on October 13. This is his view of AI-aided science, not a new release or ARC contest result. Original post read in signed-in Chrome with ordinary public engagement controls and no subscriber label; the separately read Subs feed retains its explicit private labels.
Posted .
Mazur jokes to Dan McAteer that reading his posts is a habit he will retain.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Mazur applauds Josh Wonser's account of adding graph traversal controls with a coding agent.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
In a conversation about cryptographic visualizations, Mazur says he has a similar SHA1 tool awaiting release and remains impressed by how easy these projects have become.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Mazur responds to Alexis Gallagher with a robot emoji in the coding-agent experimentation thread.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Mazur encourages Alexis Gallagher to build toward the technology he thinks could replace current products.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Mazur advises a developer unsure what to build to play and follow curiosity without immediate pressure for revenue.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Mazur says both large and small experiments can explore what coding agents can build; he says he is attempting both.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Knoop takes stock of what he sees as rapid AI disruption in coding, mathematics, games and app building, and argues that the changes create opportunities for builders.
A founder perspective on AI progress. His broad disruption and cryptography claims remain attributed commentary, separate from contest standings.
Posted .
In the DeepSeek evaluation conversation, Mazur offers to answer questions about ARC Prize testing and analysis, identifying himself as its testing lead.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Linked releaseMazur playfully encourages a commenter joking about another to-do app to think more ambitiously.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Mazur argues that developers have barely explored what current coding agents make possible and encourages experimentation.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Mazur says ARC Prize tested DeepSeek V4.1 Flash through Baseten's Models API, calculating costs at $0.30 per million input tokens and $1.20 per million output tokens.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Linked releaseARC Prize supplies reproduction, policy, leaderboard and full-result links at the end of its DeepSeek V4.1 Flash evaluation thread.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Linked releaseARC Prize links its earlier DeepSeek V4.1 Flash ARC-AGI-1/2 evaluation thread as context for its October 13 ARC-AGI-3 evaluation.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Linked releaseARC Prize’s October 13 announcement reports DeepSeek V4.1 Flash at about 3% with its provider adapter and below 1% with the standard harness. The original linked verified-results table, refreshed October 14, gives a matched Low-reasoning comparison of 3.26% versus 0.35%; its headline best configurations differ in reasoning level. Organizer model evaluation, separate from Kaggle team submissions.
Public original read without sign-in in the in-app browser; linked primary results table independently read. Organizer evaluation, separate from Kaggle team standings.
Posted .
Linked releaseMazur thanks Krzysztof Cichy for positive feedback on his project.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Knoop predicts that a future generation of game consoles will rely on cloud streaming, while quoting a report about a possible GTA 6 PC launch.
A gaming prediction and reaction to a reported rumor; neither is a verified ARC development.
Posted .
Chollet links to Kylon in the follow-up to his discussion of shared team context for coding agents.
Companion link in the same thread; group with the original discussion rather than present it as a separate announcement.
Posted .
Linked releaseChollet argues that coding-agent time savings can be erased by team coordination and review costs. He highlights Kylon, which places Claude Code and Codex in a shared team workspace.
An attributed founder view on team context and agent workflows, not measured product performance.
Posted .
Linked releaseReplying to Dries Smit’s friendly comment that Yi-Chia is giving Tufa a run for its money, Cole jokes that the competition must keep the effort level high. Parent congratulations and Smit reply read.
Light supporting color, posted October 13 afternoon and recovered during catch-up. No new score claim or evidence about actual effort, hardware or rivalry. X unauthenticated access returned 403; private pending verification.
Posted .
Smit thanks Cole and jokes that Yi-Chia is giving the team a run for its money.
Friendly banter, not a disclosure of methods or a new score.
Posted .
In an AI-risk debate, Mazur gives his view that the lab is trying to raise awareness so society can prepare. This is his interpretation of its motives.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Chollet congratulates the Gemini 4 Argon release and gives a positive first reaction to the model, quoting the release announcement.
Founder reaction to a model release; the post does not supply a new ARC score or an independent evaluation.
Posted .
Mazur proposes that withholding operational details may balance warning about risks against enabling misuse. He presents this as a possible explanation.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Japanese translation: That’s a wonderful attitude! Let’s both do our best. A supportive reply to TK256’s ARC-AGI-3 effort statement.
Public original Japanese reply readable without sign-in in root conversation. Exact Kaggle identity/team unresolved; supportive tone does not establish a roster connection.
Posted .
Bensinger advocates urgent international restrictions on large AI training runs and discusses uncertainty in dangerous-AI timelines. These are his risk assessments and policy proposals, not established predictions.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
ARC Prize announces Tufa Labs reclaimed first on the ARC-AGI-3 public Kaggle board with roughly 69%, up from roughly 56%. This is the author’s dated October 13 observation; refresh standings before publication.
Organizer confirmation of a consequential team result already covered in the newspaper’s morning/afternoon reporting; no methods disclosed.
Posted .
Mazur asks a critic why AI-danger claims might be redacted.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Japanese translation: Let’s work hard for the next three weeks! Aim for a gold medal. A supportive reply to TK256, expressing an ambition rather than an achieved medal.
Respectful competitor encouragement. Actual original Japanese reply read signed in and publicly visible in logged-out root conversation. Kaggle identity/team unresolved; no affiliation inferred.
Posted .
Kamradt says human agency, with curiosity as a component, is his strongest human competitive advantage.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Kamradt argues curiosity can differentiate people but would be a component of capable AI rather than a lasting AI advantage.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Mazur asks what evidence would distinguish an AI-risk warning from fear-based marketing for the critic.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Rieseberg introduces a tutorial on being more ambitious with AI models. The linked October 13 essay distinguishes higher-level goals, deeper professional methods and cross-field approaches, with outside validation. Matt Mazur’s repost was first observed this shift; its action date is unknown.
General AI-use background captured before newspaper selection; no new ARC score or competition method is disclosed.
Posted .
Linked releaseKamradt argues that compute energy and data are lasting competitive advantages and that money can increasingly purchase knowledge. This is general AI commentary.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Mazur warmly acknowledges Val Sopi appearing in a project-update video.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Japanese translation: ARC-AGI-3 has become my highest priority. Compared with my teammates, I lack knowledge and experience, so I need to spend more time finding improvements and work hard enough that they feel glad they included me.
Own competitor voice, original Japanese and exact date publicly readable without sign-in. Owned site links X and Kaggle; Kaggle links back to X. Existing dated roster lists tk256ailab on ARC-3 team 16569113; no joining date or responsibility for a score inferred.
Posted .
In a discussion of automated engagement, Mazur says a large portion of the replies he receives appear similar.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Mazur responds with an agreement emoji to Boardy's encouragement to push a project until it breaks.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Mazur says Scene Studio would have been difficult for him to build as well without a capable coding agent, describing it as a way to test the agent's limits.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Scene Studio example: Mars in post-painterly abstraction style; Mazur links the tool and describes its generation workflow.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Linked releaseScene Studio example: Paris and the Eiffel Tower rendered as an etching.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Linked releaseScene Studio example: a hilltop castle in redwork embroidery style.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Linked releaseScene Studio example: a futuristic city in pixel art.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Linked releaseScene Studio example: a fireplace library in midcentury travel-poster style.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Linked releaseScene Studio example: a tropical sunset beach in enamel grisaille style.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Linked releaseScene Studio example: a lunar scene with Earth overhead, rendered in Amiga HAM pixel style.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Linked releaseMazur introduces his Scene Studio project, which he says he built with Opus 5.5 to generate animated scenes in 100 styles, with GIF downloads, sharing and favorites.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Linked releaseKamradt comments on competition and profit margins while quoting a claim about AI sales agents and declining email response rates. The quoted statistics are not independently established here.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Replying to the Digest’s morning announcement, Jean-François Puget asks why he was singled out and emphasizes that NVARC is a team with other members.
A direct, public request for team credit supports a short roster profile rather than treating a team result as an individual accomplishment.
Posted .
Mazur highlights a warning about increasingly capable AI replicators while quoting Andrew Curran. This is commentary on a reported risk, not a demonstrated contest result.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Chollet jokes about agriculture as self-replicating, recursively improving technology, using the multiplication of seeds as his example.
Light founder commentary, posted on the evening of October 12 Eastern; it is not an AI research result.
Posted .
Chris summarizes an attributed Anthropic warning about models replicating on heterogeneous consumer hardware. His post relays reported experiments and forecasts; the underlying research was not independently verified in this scan.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Zeng links the Schema preprint, public implementation repository and project site, and credits the research team. The repository was read without sign-in.
Direct primary-source links for reporters. Code was inspected as source material, not run or installed.
Posted .
Linked releaseIn an author-reported ten-public-game ablation, removing executable modeling, history checks, planning or runtime verification reduces performance.
Explains the design claim; grouped with the release rather than a separate score story.
Posted .
Linked releaseZeng reports that Schema with Claude Fable 5 clears the 25 public ARC-AGI-3 games and improves relative human action efficiency over Claude Code. The paper supports this author-reported public-set comparison.
Useful technical context for the release; the public tutorial set does not establish competition performance.
Posted .
Linked releaseZeng announces Schema’s full report and code, building on its July public-game demonstration. The linked preprint was submitted September 30; this is an October 12 release announcement.
A fresh release lead with publicly accessible paper and implementation. Public-game evaluation; no private Kaggle result established.
Posted .
Linked releaseEnglish translation of the Japanese post: fshindo thinks local LLM enthusiasts with many GPUs who improve harnesses and inference speed ought to enter competitions, then jokes about discovering formidable people who would become rivals. This is a friendly invitation, not a claim about a named rival’s resources.
Japanese competitor voice about the connection between practical inference/harness work and contest participation.
Posted .
English translation of the Japanese post: fshindo says they finally reached the medal range in ARC-AGI-2 and are pleased because that contest had received the most of their time and effort without results until then. The profile ffsscf directly links this X account; the claim is dated to the post, not a current rank.
A useful participant account of persistence and uneven progress between the two contests.
Posted .
The author asks in Japanese whether this counts as a good result, then shares Kaggle’s rank-31 card for ARC-AGI-2. The linked competition is verified; their exact Kaggle account/team remains unresolved.
Additional Japanese competitor voice without guessing a present board identity from an old rank.
Posted .
Malysheva says most of the Mostik team is together on site for brainstorming and that Stanislav Smirnov joined. The post shows ongoing activity but announces no new ARC submission, score or method.
Current activity helps avoid treating a low board position as evidence that the company abandoned the race.
Posted .
Mazur jokes that the reason a model-evaluation chart has not been updated could be that the tests are still running. This is a joke, not an explanation from METR.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
PunkCyber tells the ARC Daily Digest that 復活の混テキスト is their team. The exact Japanese name matches entries on the current ARC-3 and ARC-2 boards; these have separate competition team IDs.
A competitor checked in with the paper and supplied a direct team identification. It supports a friendly shout-out; the reply does not identify their individual Kaggle handle or establish who produced a submission.
Posted .
Posting in Japanese, fshindo said the ARC-AGI-3 competition had taught them a great deal about agent harnesses and called contests a valuable learning opportunity. The post links the current Kaggle competition.
A first-person ARC-3 engineering account, now connected through the direct Kaggle profile link to team 16992721. It is an attributed experience, not a claimed cause of any score.
Posted .
Replying to Michael Taiwo’s suggestion of a small-budget version students could run before attempting the full experiment, Chollet called it a great idea. The conversation concerns his previously announced KerasHub/JAX/TPU post-training experiments; it does not announce a release, schedule or ARC score.
A modest follow-up to the archived post-training lead: student accessibility is a possible direction welcomed by Chollet, not a promised deliverable or evidence of a competition method.
Posted .
Mike Knoop said the same thing is happening with ARC-AGI-3 while quoting Pedro Domingos’s comment that models are absorbing harness behavior. The quoted post links Harness Learning Enables Generalizable Test-Time Adaptation. Knoop gives no named team, method, score or ARC experiment. The linked paper trains a proposer to revise executable harness code using feedback; its reported evaluations cover reasoning and multi-hop question answering, so the ARC-3 observation remains Knoop’s attributed comment.
Fresh ARC co-founder reaction for reporting on how models and harnesses divide the work. Useful research context, with no new verified Kaggle result or evidence that harnesses disappear.
Posted .
Linked releaseMazur jokes that a model describing a task as fun is a sign of sufficient ambition.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Pedro Domingos described a linked paper as showing the model absorbing the harness. The primary release, submitted September 28, describes learning to revise executable harnesses from execution feedback, with reasoning and multi-hop question-answering evaluations. Its introduction retains an executable harness and frozen models at test time. The social phrasing is Domingos’s interpretation; it should not be reported as a demonstrated removal of the harness or an ARC competition result.
Primary quoted context for Mike Knoop’s new ARC-AGI-3 reaction. Group by the shared paper and distinguish its actual experiment from the broader social interpretation.
Posted .
Linked releaseFeiZ attributed the one-level difference to context compression settings in a follow-up. A reply questioned the comparison setup; the explanation is an author claim, not independently established.
Necessary context and caveat for the root ARC-AGI-3 experiment.
Posted .
FeiZ reported a one-level ls20 ARC-AGI-3 comparison: a DS Bot cleared that level within 81 steps, while a DeepSeek Harness exhausted 110 steps without clearing it. This is the author’s own small test, not a Kaggle standing or verified benchmark result.
Recent ARC-AGI-3 community experiment; the author explicitly says one level proves little.
Posted .
The author argues that current model-written software is strong enough to outgrow the term vibe coding. This is an opinion about coding quality.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Takuma Mori (@menphim) shared an October 11 ARC-AGI-3 card claiming rank 76, with Kaggle’s joking addiction line. Direct Kaggle-to-X and owned Takuma Mori page-to-Kaggle links now verify the identity. This is the dated claim, not a current ranking; team observations are in the separate identity handoff.
A dated standings lead outside the existing watch list; no current ranking or method is inferred.
Posted .
masspeaks reacted in Japanese to Tufa Labs exceeding 85% on ARC-AGI-2, calling the result impressive. The public board independently shows Tufa Labs at 88.06% in the afternoon snapshot.
Respectful reaction from a current ARC-AGI-2 mover to the leader.
Posted .
Linked releaseIn a Japanese post, masspeaks said careful preparation had paid off and expected a high gold-band cutoff. This is the contestant’s reaction; the post alone does not establish the exact score or method.
Context from the ARC-AGI-2 public-board mover; compare any score or rank with the board brief.
Posted .
The author describes AI-generated mathematics as unfamiliar and predicts similar changes in other sciences. These are the author’s claims and forecasts, not independently verified discoveries.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Replying to a question about TPU sharding for his LLM post-training experiments, Chollet said the keras.distribution API takes care of it. This is a clarification of his earlier experiment note, not a release announcement.
Adds a concrete implementation detail to the previously archived research lead; follow up if the experiments are released.
Posted .
Chollet said he has been trying LLM post-training experiments with KerasHub, JAX and TPUs, and might open-source some of the work. He did not announce a release or an ARC competition result.
A research-direction note from an ARC Prize co-founder; revisit if he publishes the experiments.
Posted .
Mazur says the coding agent produced the cartoon lip-sync implementation correctly on its first attempt, in his experience.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Barney joked that he had reached the top ten in the ARC-AGI-3 Kaggle competition, linked the competition page and a video, and called himself a vibe coder. This is his own claim, not an independently checked standing.
A current contender’s lighthearted reaction to a claimed top-ten position; verify the dated board before using the rank as fact.
Posted .
Linked releaseMazur asks Fraser Price to speculate about what might eventually limit evolution in the simulation.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Mazur asks about the upper bound of evolution in Fraser Price's virtual environment.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Mazur compliments Mark Barney's AI music experiment.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Mazur asks Mark Barney how the ARC Daily Digest is generated and how much of the process is automated, explaining that he is curious.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Mazur says many replies appear AI-generated but explicitly remains unsure whether this commenter's replies are; he says they are decent.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Mazur conditionally compliments the quality of a commenter's replies if they are AI-generated, without establishing their origin.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Mazur thanks Hasnain for sharing the cartoon project.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Mazur suggests it could be fun to turn the cartoon concept into an interactive branching-adventure game.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Mazur compliments RidgetopAi's game preview.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Mazur says the agent wrote the lip-sync logic for every cartoon.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Chollet shared a chart and argued that AI shifted from token-by-token answer completion toward synthesizing reasoning chains on the fly, which he says helped coding agents work.
A founder’s view of the shift in agent reasoning offers context for ARC-AGI-3 coverage; this post reports no competition result. The individual post was readable without sign-in.
Posted .
Kamradt confirms his city prototype uses each city's actual sunset timing.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Mazur acknowledges imperfections in the cartoon output but calls it a strong first attempt compared with what he could produce himself.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Mazur replies with an agreement emoji to Ross Morsali's enthusiasm for Final Fantasy.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Kamradt shares his Pixie Citie prototype, describing city views connected to map data, live weather, San Francisco trees and music.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Knoop suggests three ways to help AI agents discover and use a product API: awareness through training or search, publishing evaluation scores, and training for the required skill.
Practical founder advice from October 10, recovered during catch-up; do not relabel it as an October 13 announcement.
Posted .
OCEAN highlighted the Memento 3 preprint. Its authors report that a fixed language model revises an external rulebook and executable world model; on the 25 public ARC-AGI-3 games, they report clearing every level with fewer actions than the human baseline. These are author-reported results on the public set.
A new ARC-AGI-3 research lead for the edition desk. The paper provides a concrete mechanism and results, but the public tutorial set cannot establish private-game performance.
Posted .
Linked releaseEnglish translation of the Japanese post: masspeaks says submissions older than 48 hours can disappear from Active Events while still scoring, and jokes about discovering that none of their submissions had finished. This is a participant report, not an organizer-confirmed platform rule.
Explains why some public competitor chatter concerns scoring waits rather than a new method or result.
Posted .
Cartoon example: a missing-sock movie trailer interrupted by a parent, in VHS anime style.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Cartoon example: a historical radio announcer discovers a small podcast audience, in Art Deco style.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Cartoon example: an aristocrat shares tea with a cowgirl, in claymation style.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Cartoon example: a cab driver encounters a patient turtle, in pop-art comic style.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Cartoon example: an influencer judges a wizard's outfit, in illuminated-manuscript style.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Cartoon example: a nature narrator observes furniture assembly, in field-guide watercolor style.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Cartoon example: a vampire attends a sunrise meditation class, styled as a silent film.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Mazur presents eight short dialogue cartoons in different styles, credits ElevenLabs voices and GPT Image 2 art, and says Opus 5.5 wrote the animation, timing, captions and lip sync. He shares a reproducible workflow; this is his project account.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
In a reply to Mike Knoop’s ARC-AGI-2 bonus-threshold post, Tufa Labs said it had not set out to solve ARC-AGI-2, started by accident, and plans to share more after the deadline. This is the team’s account, not an independently established history.
First-person context from the ARC-AGI-2 public-board leader; no method or final result is disclosed.
Posted .
Linked releaseMazur asks what controls the physics in Fraser Price's simulation.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Mazur playfully celebrates positive feedback on the orb visualization.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Mazur compliments Nathan Helm-Burger's orb visualization and asks what it does.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Mazur responds with mock alarm to a joke that he has become a pending tool call.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Mazur praises the coding agent's work in his bug-fix anecdote.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Mazur links his earlier account of letting Claude Code work continuously on his SaaS product; the linked experiment dates to September.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Mazur says he permits the agent to merge and deploy work within his authorized scope without prior PR review, and reports no problems so far. This is a personal experience, not a general safety finding.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Mazur shares a view of the pull request supporting his bug-fix anecdote.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Mazur says he withheld some provocative image-grid experiments after they exposed model bias he did not want to amplify.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Mazur clarifies that the database repair script identified affected records when executed rather than relying on a fixed list of record IDs.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Mazur says his autonomous coding agent found an old SaaS bug, deployed a fix and prepared a repair script because it lacked production database write permission. He says he later ran that script himself and reflects on being the human bottleneck.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Mazur points to an older diversity-themed image-grid experiment in a discussion of visual storytelling.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Mazur points to his April experiment depicting stages of an ancient Roman person's life in a text-free image grid.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Puget asked ARC Prize whether the 85% bonus threshold refers to the public or private leaderboard; no organizer answer was visible in the checked thread.
Flags a reporting question; avoid presenting the public score as the final prize result.
Posted .
Linked releaseMalysheva discusses Vision Wormhole’s use of vision tokens as a continuous communication interface between models and highlights losses on harder tasks with larger senders. The desk opened the authors’ arXiv abstract and owned project page; the work is about cooperative VLM communication, not an ARC leaderboard result.
The size of a reasoning handoff may constrain what the receiving model can preserve.
Posted .
Linked releaseMazur applauds Mark Barney's report that a curation and polish pass had expanded in scope.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Mazur says he had not known the ARC Daily Digest site existed and compliments it.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Cole congratulated Tufa Labs on the ARC-2 result, calling it a momentous achievement.
A named competitor’s respectful reaction adds color without implying rivalry or shared methods.
Posted .
Linked releaseKnoop said ARC-AGI-2 remains much harder than ARC-AGI-1 and that eventual open solutions will be useful to study, while quoting his earlier reaction to Tufa Labs crossing the public Kaggle bonus threshold.
Organizer context about the contest; his forward-looking statement is not a released solution or final award.
Posted .
Linked releaseKnoop said the 85% Grand Prize bonus threshold had been reached on Kaggle and called 2026 the final ARC-AGI-2 Kaggle year.
Provides organizer context for the Tufa Labs result; his expectation about future open artifacts is not yet a release.
Posted .
Linked releaseChollet reacted to ARC Prize’s Tufa Labs ARC-2 announcement by saying Kaggle scores are getting very good.
A founder reaction adds a measured voice to the organizer’s score announcement.
Posted .
Linked releaseARC Prize announced a new 2026 ARC-AGI-2 Kaggle high of 88.06% by Tufa Labs and said a $150,000 bonus would be split among teams scoring above 85%.
A public-board threshold announcement and major standings lead; the post does not establish a final private score.
Posted .
Linked releaseARC Prize announced Yi-Chia Chen at 59.17% on the 2026 ARC-AGI-3 Kaggle board, taking the lead over Tufa Labs.
Historical October 9 standings announcement, newly inventoried as a Chollet repost; not a fresh score move. Original and visible context re-read; existing independently verified public visibility preserved.
Posted .
Linked releaseKamradt said he was excited Wolfram would join the ARC Prize Summit in two weeks, quoting the organizer’s announcement.
Confirms the organizer’s program announcement from the ARC Prize president.
Posted .
Linked releaseARC Prize announced Stephen Wolfram as a keynote speaker for its 2026 Research Summit.
A current event-program lead for the meeting digest.
Posted .
Linked releaseMazur applauds Dave Martin's effort to share more on X.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Tufa Labs made a tongue-in-cheek post about a temporary setback and “overwhelming resources and compute”; the post included AI-marked imagery.
Competition banter only; the line is not evidence of the team’s actual budget, compute or methods.
Posted .
AadEd shares an October 8 ARC-AGI-3 rank-590 claim and attributes the improvement to small prompt changes, alongside a RAM joke. Exact Kaggle account/team remains unresolved.
Older catch-up contender discovery, not an overnight gain or verified hardware/method claim.
Posted .
Linked releaseMalysheva highlights CacheFuse as a training-free blending of compatible Qwen3-8B and Qwen3-4B caches. She stresses that matching cache shapes and a shared tokenizer are required and contrasts this with cross-family bridges. Her reported numerical comparisons are attributed to the paper; this desk has not independently read the CacheFuse paper.
A concrete reading lead about reusable inference state and compatibility limits, not a verified ARC score.
Posted .
Kamradt shared images and called Ouellette’s release a “V3 inspired dataset.”
An organizer reaction to a community resource; the original release and repository hold the substantive details.
Posted .
Linked releaseOuellette released an ARC-AGI-3-style training-data framework with 429 playable games, solvers and demonstration generators; he said 37 games still lacked solvers. The linked repository says it combines augmented public ARC demos and other game sources.
A concrete research resource; its author describes training material, not a measured Kaggle-score gain or coverage of private games.
Posted .
Linked releaseMazur asks how Fraser Price would distinguish life from an interesting visualization in his simulation.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Mazur contrasts narrow requests for unit tests with asking an agent to inspect and improve tests throughout a codebase, as an example of adjusting ambitions.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Translation of the original Japanese: masspeaks asks for help with an ARC-2 queue that takes a long time and ends in an error, while describing one submission per day. This is the author’s October 8 complaint, not an organizer ruling or evidence of a current queue outage.
Older catch-up context for queue reporting; current organizer decisions must come from the original organizer thread.
Posted .
In a reply to a question about AI subscriptions and GPUs, masspeaks says they have been using Claude Code and some A100 GPUs, and jokes about wanting unlimited machine resources. The conversation sits beneath their ARC-2 start post.
A narrow first-person resource disclosure, not a full Kaggle recipe, spending figure or current hardware inventory. Older catch-up material.
Posted .
Mazur responds with a handshake emoji to Mark Barney's offer to coordinate after curating ARC resources.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Translation of the original Japanese: masspeaks says they have started ARC-AGI-2. The thread contains friendly greetings from another participant. This is an October 8 catch-up source, not a fresh entry announcement.
Dated Japanese competitor voice for the desk; no exact team identity or current rank established.
Posted .
In the discussion of his temperature-zero determinism post, Jean-François Puget says not to take the hiring part too seriously: he used it to draw attention to the topic.
Context for the Oct 7 inference reproducibility lead; prevents treating the rhetorical hiring remark as an actual policy announcement.
Posted .
Mazur replies to Dave Martin's career reflection with a GIF. Its content was not interpreted; no further meaning is asserted.
Priority-account capture before editorial selection. Original read in signed-in Chrome; separately readable without sign-in. General commentary is not a competition result.
Posted .
Quantum Howl shares an October 7 ARC-AGI-3 rank-175 Kaggle card. Exact Kaggle account and team identity remain unresolved.
Older catch-up contender discovery; neither a current standing nor a verified identity mapping.
Posted .
Linked releaseJean-François Puget warns that setting temperature to zero does not guarantee fully deterministic LLM inference. The post quotes a claim that temperature zero always produces the same answer. In a later reply he says the hiring wording should not be taken too seriously and was intended to draw attention to the topic.
Older catch-up lead for reporting on reproducibility and run-to-run variation. This is Puget’s general inference comment, not a measurement explaining any ARC leaderboard result. Preserve the clarification when using the hiring line.
Posted .
ARC Prize published verified Grok 4.7 scores: 1.8% on ARC-AGI-3 with its standard harness, 10.0% with a provider adapter, and 61.4% on ARC-AGI-2.
The harness distinction matters when comparing model scores; the post reports organizer verification.
Posted .
The author says he has observed quantitative and technical firms hiring thinkers and generalists, and predicts acceleration. Links an AQR essay on creativity and emotional intelligence. Full original and visible context read; logged-out public body confirmed. Older newly inventoried Matt repost; action time unknown.
General AI commentary captured before selection. Prediction remains attributed; no ARC affiliation or competition recipe inferred.
Posted .
Linked releaseARC Prize published verified DeepSeek V4.1 Flash scores of 72.9% on ARC-AGI-2 and 94.5% on ARC-AGI-1, with stated per-task costs.
A dated organizer model verification relevant to ARC-2 comparisons.
Posted .
Kamradt invited researchers to contact ARC Prize about red-teaming future ARC-AGI benchmarks and mentioned grants and credits as possible support.
A direct research-participation lead; it is an invitation, not a claim about any current contestant’s approach.
Posted .
Replying to Alex Zverianskii, Malysheva distinguishes fine-tuning, which puts capability into the small model’s weights, from a learned bridge that lets the small model use a larger model’s computation during inference.
Clarifies the proposed inference-time relationship; it is not a disclosure of Mostik’s Kaggle submission.
Posted .
Malysheva proposes testing a domain-fine-tuned small receiver with a newer large sender and a newly trained bridge. Preserving the receiver’s tuning through the sender swap is the experiment she wants to run, not a completed demonstrated result.
Research lead on whether specialist tuning can survive changes to a reasoning supplier.
Posted .
Swiz shared a native Kaggle ARC-AGI-2 rank-14 claim on October 5 while requesting model-training credits. This is an older posting-time claim; exact Kaggle identity and team remain unresolved.
Newly discovered older competitor voice beyond the watch list. Original body and timestamp read in signed-in Chrome and independently without sign-in. No current rank, budget or team identity inferred.
Posted .
Linked releaseARC Prize announces the three qualifying open-source ARC-AGI-3 Milestone 2 winners: Daniel Franzen, Lord Han Solo and Lohit Siriki. Milestone prize placement is distinct from the unrestricted ongoing public leaderboard.
Primary historical result for the former-leaders morning feature.
Posted .
Liang describes a playable explanation of ARC-AGI-3 made in collaboration with ARC Prize, linking public tasks and Astra replay notes. His September 30 original quotes Greg Kamradt. Original and context read; logged-out public body confirmed. Older newly inventoried Greg repost; action time unknown.
Background teaching resource; primary page distinguishes provider-adapter and standard runs. No new Kaggle result inferred.
Posted .
Linked releaseARC-AGI-3 · @ishankarnick
New reply: Mirza Milan Farabi reports slow dataset/model downloads and model precaching/loading taking two or three times as long; says a run scored zero and ended after about thirty minutes. Earlier replies from Jeroen Cottaar and CPMP discuss the fifteen-minute no-move timeout and slow loading. These are participant reports, not a new host ruling or confirmed outage.
Activity · 5 replies
ARC-AGI-2 · @anlthms
mikelou1 reports an 87% local Qwen run under time constraints at roughly four-L4 speed (October 20 03:00 UTC). At 06:26 UTC the same author explicitly says 87% likely will not transfer, would be surprised by even 80%, and hopes for roughly 75%; those are expectations, not a submitted result. A later reply at 08:54 UTC says they are not very concerned about overfit because they get few runs. Other participants discuss throughput and the local-to-leaderboard gap. No complete recipe or independently verified new leaderboard score was supplied.
Activity · 16 replies
ARC-AGI-3 · @ruichardliu
Expanded the newest nine-reply branch. Son Pham’s October 19 23:58 UTC reply encourages a participant after a disappointing submission, recounting his own overconfidence in public evaluation and emphasizing completely unseen testing. At October 20 05:26 UTC, CPMP says an earlier ARC-AGI-1 solver went from 75% evaluation to 6% on submission, before his team’s subsequent ARC-AGI-2 win. These are attributed recollections, not new contest scores. Deleted posts remain deleted; a third-party score allegation was not promoted as a verified result.
Activity · 77 replies
ARC-AGI-3 · @franciseliang
Francise Liang reports stopping work and shares tests of Franzen’s public Milestone 2 notebook on an RTX Pro 6000, with 25 public games and two passes per run. Two unchanged runs averaged 47.30 and 48.92; per-game differences had a reported standard deviation of 24 points. Increasing concurrent streams from ten to fourteen reduced the reported mean from 47.30 to 29.90 and throughput; stopping stalled levels at 110k tokens gave 38.97 versus a two-run baseline average of 48.11. Liang recommends repeated comparisons. These are participant measurements on public games, not independently reproduced results or hidden-board scores. Reply author xiaoz259 attributes poor throughput to insufficient KV-cache memory and lane preemption.
Activity · 1 replies