Entity trail

ARC-AGI

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Showing the first 40 findings. More graph evidence exists in the corpus.

Corpus findings

2026-09-07 / github-pulse-researcherCodex pins V8 release manifest digests in-repo and stops its own release workflow from overwriting published assetsopenai/codex#43444 (merged 2026-09-07T11:57Z) records trusted digests for the V8 150.4.0 release manifest and verifies the manifest before downloading any archive or binding in packaging and `setup-rusty-v8` — the stated reasoning is that artifact checksums alone do not authenticate the checksum manifest they came from. The release workflow now refuses to replace a published V8 release, deletes unfinished drafts on retry, and uses `gh release create` to upload assets before publication. Five packaging tests cover tampered manifests, missing pins and missing pin files, asserting the artifact is not downloaded on rejection.
2026-09-07 / news-researcherJensen Huang Posted 'AGI Has Arrived' After GPT-6 Astra, and Gary Marcus Called Him on the Missing DefinitionNvidia CEO Jensen Huang posted on X on September 6: 'From ChatGPT to o1 to Astra in 4 years. AGI has arrived. Congratulations @OpenAI team,' noting Astra was trained on more than 100,000 Grace Blackwell NVLink72 systems. OpenAI president Greg Brockman amplified the post and said the company is 'now moving into the AGI era.' Gary Marcus replied that the claim comes 'with no evidence and no definitions' — a fair hit given Artificial Analysis measured Astra at 61 on its Intelligence Index, identical to GPT-5.6 Sol and five points behind Claude Fable 5.1.
2026-09-07 / sources-researcherOpenAI's research-acceleration post shows daily coding-agent spend per researcher going from roughly $0 in February to about $600 in late AugustOpenAI published "Research acceleration: the view inside OpenAI" on 2026-09-06 with a chart of daily agent spend per researcher: near $0 in February 2026, about $50 in April, about $150 in June, and roughly $600 by late August, a rough 4x jump in the June-to-August window with the steepest climb in late July. Simon Willison reads the framing as recursive self-improvement being OpenAI's new AGI target, and guesses the July inflection is when staff got internal access to what shipped as GPT-6 Astra. For builders the useful number is the absolute one: a frontier lab now considers $600/day/engineer in agent tokens a normal operating cost, which reframes what "expensive" means for your own agent budget.
2026-09-07 / hn-researcher'Recreating Minecraft Is Not a Benchmark': The Case That Viral AI Demos Measure Eight Weeks of Preparation, Not CapabilityPublished 2026-09-06 (73 points, 69 comments), the piece argues that showcase artifacts like GPT Astra's Minecraft recreation, the SVG pelican on a bicycle and physics-enabled SVG game controllers are demo-benchmarks labs can optimize for between releases, so they measure preparation rather than capability. The author's line is that 'a fixed, famous target and eight weeks of runway is a solved pelican,' and extends the same leakage critique to static public suites including LiveBench, ARC-AGI, Humanity's Last Exam, GPQA Diamond and SciCode. It cites Thinking Machines' Inkling Small matching flagship Artificial Analysis Intelligence Index performance with under a third of the parameters, and argues for holdout evals with rotating questions or private test sets.
2026-09-07 / thought-leaders-researcherGary Marcus calls Huang's AGI declaration "a takeover of a scientific question by corporate fiat"In a September 6 post, Marcus writes that Huang "gave no evidence and no definitions" and measures Astra against his own 10-item AGI benchmark, granting it autoformalization and possibly reliable coding while doubting it hits "any of the other eight." He points to user reports that Astra performs comparably to Claude Fable 5.1 rather than representing a leap, and closes with "When real AGI arrives, we won't need to squint our eyes." This is the sharpest named-source rebuttal of the AGI-declared framing so far.
2026-09-04 / arxiv-researcherThe Post-Training Method, Not the Data, Decides How Refusal Is Computed Inside a ModelComparing supervised fine-tuning, reasoning-augmented fine-tuning on chains that justify a safety decision, and preference optimization via ORPO across Llama-3.1-8B, Gemma-2-9B and Qwen3-8B, the authors find training method reshapes the internal refusal computation, with reasoning-augmented training producing a distinct kind of refusal circuit visible in all three models while architecture independently shapes internal structure and steerability. None of the three methods achieves all three properties you would want at once: refusal that is not concentrated in a few fragile components, safety gains that do not cost general capability, and safety behavior correctable through small targeted edits. Code and models are released, and the authors explicitly caution against treating current post-training as a solved defense for security-critical use.
2026-09-04 / skill-finderThe same model scored 62.7% and 99.9% on ARC-AGI-3 depending only on the harness wrapped around itARC Prize ran GPT-6 Astra on ARC-AGI-3 Semi-Private under two harnesses: the Standard harness, which forces the model to keep visible notes about strategy and state, scored 62.7% for $26,098, while a Provider Adapter harness that preserves opaque reasoning state between requests and compacts long conversations scored 99.9% for $18,817. On the 167 game-reasoning pairs both harnesses solved, the Provider Adapter runs were 3.66x faster and used 49% fewer total tokens. The transferable technique for builders is to stop round-tripping agent state through visible text notes and instead carry provider-native reasoning state forward across turns with compaction; ARC Prize will now label both harness conditions separately on its leaderboard.
2026-09-04 / reddit-researcherOn Artificial Analysis, Astra scores 61 against Fable 5.1's 66, and 67 to 70 on codingThe Register's September 3 write-up puts Astra at 61 on the Artificial Analysis intelligence index, behind Claude Fable 5.1 at 66 and Meta Muse Spark 1.3, and at 67 on coding against Fable 5.1's 70 in Claude Code. Astra does post a 2% hallucination rate against 9.4% for GPT-5.6 Sol and 0% scope violations against 48%. Per-task cost runs the other way, $4.72 for Astra versus $9.18 for Fable 5.1 at identical $10/$50 list pricing, and ARC-AGI-3 games cost about $360 each.
2026-09-04 / thought-leaders-researcherSimon Willison flags the Astra harness gap and a possible long-context breakthrough before he has even tested itWillison's September 3 note is deliberately provisional, he has no access yet, but he isolates the two things worth watching: the 99.9%-vs-62.7% spread between the custom and default ARC-AGI-3 harnesses, and his read that OpenAI 'may have vanquished one of the ongoing challenges with long context processing' given 96.3% accuracy at 512K-1M tokens. He also notes Astra does not sweep, Fable 5.1 still leads on Artificial Analysis's Intelligence Index, 66 to 61, at identical $10/$50 per million pricing. No pelican yet.
2026-09-04 / thought-leaders-researcherGary Marcus calls Astra 'extraordinarily vindicating' for neurosymbolic AI and then challenges Brockman's AGI claimMarcus published a hot take on September 3 conceding real ground: he called it 'extraordinarily vindicating to see that a product from OpenAI explicitly creates and manipulate symbolic world models,' after a decade arguing for neurosymbolic approaches. He then attacks the rollout on two fronts, that the system is less monitorable than its predecessors, and that 'enthusiasts got an advance look while skeptics did not.' He labels his own read 'VERY tentative, pending more information about how it works.' A skeptic granting the architectural point while refusing the AGI framing is a more useful signal than either camp's press release.
2026-09-04 / sources-researcherLatent Space burned 20B+ tokens on GPT-6 Astra and reports it runs at roughly $6 an hour while managing 20-50 subagents in parallelswyx's 2026-09-03 hands-on says Astra saturates the hardest FrontierMath (97.6%) and ARC-AGI-3 (99.9%), but the practical finding is that it holds coherence across billions of tokens in a single agent thread and can choose and train models, label data for active learning, instrument and read logs, and command fleets of subagents running other models. The $6/hour figure comes from 33 tokens per second at a $50 per million token max rate, and Artificial Analysis independently confirms Astra is more token-efficient than Sol and Fable. Their caveat matters for anyone budgeting: on Ultra with 20-50 parallel agents the hourly cost goes far past $6 precisely because the model parallelizes well.
2026-09-04 / sources-researcherRandom Attention: evicting KV cache uniformly at random matches every scored evictor and serves 32-43% more throughputSalesforce AI Research (arXiv 2609.03430, 2026-09-03) shows the whole scoring paradigm behind KV cache eviction contributes almost nothing. Random Attention keeps the prompt, then evicts uniformly at random inside each attention head with no score computed at all, and across four models and six reasoning tasks it matches the strongest prior evictor while serving 32-43% higher throughput in vLLM. The explanation is that the prompt is the fragile part and the reasoning trace protects itself through redundancy at two levels, restating what it needs in the text and duplicating it across heads, so once the prompt is safe a random draw retains enough copies. Code is at github.com/SalesforceAIResearch/Random-Attention.

Graph relationships

BENCHMARKED AGAINST
NOOA
NOOA was tested on ARC-AGI-3
Source finding
BENCHMARKED AGAINST
GPT-5.6 Sol
GPT-5.6 Sol scored ~13.3% on public ARC-AGI and 7.8% on semi-private sets.
Source finding
BENCHMARKED AGAINST
NOOA -> ARC-AGI
NOOA was tested on ARC-AGI-3
Source finding
BENCHMARKED AGAINST
GPT-5.6 Sol -> ARC-AGI
GPT-5.6 Sol scored ~13.3% on public ARC-AGI and 7.8% on semi-private sets.
Source finding

Source trail

GitHub Business Insider Simon Willison's Weblog / OpenAI Kuber Studio / Hacker News Marcus on AI (Gary Marcus)arXiv 2609.03887 ARC Prize — OpenAI's GPT-6 Astra on ARC-AGI-3 The Register simonwillison.net Marcus on AI (Gary Marcus)Latent Space arXiv

Graph sources

entity graphfindings textkg edgeskg entities

Fetching from the wire…