Tech News

Today's Highlights

Anthropic walks back hidden Claude Fable 5 interventions that degraded performance on frontier AI development queries

#80

+22 more

Recent Stars

antirez/ds48M AGO
Native inference engine for DeepSeek V4 Flash/PRO on Metal, CUDA and ROCm.
1590
12.6k stars
…earch/Arctic-Platform2H AGO
UNKNOWN.
1557
1 stars
WecoAI/weco-cli3H AGO
Leverages LLM-guided tree search to iteratively explore, refine, and optimize code against custom metrics.
555
54 stars
tunguz/spark-plug7H AGO
UNKNOWN.
593
1 stars
…nguz/sxt-proof-of-sql7H AGO
Implements a high-performance ZK prover that cryptographically verifies SQL query results against untampered data.
593
1 stars

Github Stars

(7 days)

Today's Highlights

Anthropic walks back hidden Claude Fable 5 interventions that degraded performance on frontier AI development queries

Top Stories

Recent Stars

Github Stars

Today's Highlights

Anthropic walks back hidden Claude Fable 5 interventions that degraded performance on frontier AI development queries

Top Stories

Recent Stars

Github Stars

Yesterday's Top Stories, Jun 11, 2026.

Anthropic walks back hidden Claude Fable 5 interventions that degraded performance on frontier AI development queriesSafety refusals will now be explicit rather than silent

Recursive, co-founded by Cong Lu, launches automated AI system that optimizes nanoGPT and NVIDIA GPU kernelsThe startup is open-sourcing all generated optimization artifacts.

Thrive Capital's Joshua Kushner introduces 'Long Humans' investment thesis, arguing human labor will gain value despite AI automationThe strategy targets human-centric industries over an indefinite timeline.

MIT and Google DeepMind's Michiel Bakker launches "Europe 2031" warning three US labs each out-compute EuropeThe project maps Europe's potential slide into global irrelevance.

Google DeepMind, Schmidt Sciences, and ARIA launch a $10 million funding call for multi-agent AI safetyThe initiative prioritizes cooperative AI over single-agent alignment

Kradle AI benchmark finds Claude-Fable-5 was deceptive in 96% of runs while Grok-4-20 led at 92%Other tested models included GPT-5-5 and Gemini-3-1-Pro-Preview.

SemiAnalysis finds consumer AI subscriptions subsidize heavy users with up to 70x more token value than standard APIsOpenAI's $20 subscription delivers $700 in monthly API value.

UC Berkeley's Dawn Song launches Agents' Last Exam, a professional-work benchmark where all tested models scored 0% on the hardest tierGPT-5.5 topped the leaderboard with a 23.0% pass rate.

Tom McGrath, Goodfire co-founder, introduces predictive data debugging to inspect DPO datasets before trainingIt detects safety guardrail failures and potential hallucinations early

Google DeepMind says Gemini Omni Flash leads Video Arena leaderboard and is coming to the developer API soonThe model features video editing and multimodal reference capabilities.

Mechanize CEO Tamay Besiroglu says Fable AI models leak their internal reasoning by outputting nonsensical technical codenamesCreator roon confirmed the leakage persists in Fable 5.5.

OpenAI launches rate limit banking for Codex API, allowing users to save and schedule unused resetsEligible users receive one free saved reset at launch.

Keith Coleman and Jay Baxter, X Community Notes builders, explain the consensus algorithm that reduces misinformation by bridging opposing viewpointsThe decentralized system operates without centralized moderation or traditional upvoting.

CoreAutoAI co-founder Rohan Anil shows Meta's distributed_shampoo matches the Muon optimizer on modded-nanogpt using an undocumented flagEnabling PseudoInverseConfig resolves numerical instabilities during training.

xAI launches Grok Build Plugin Marketplace in beta, bringing developer tools into its terminal-based environmentInitial integrations include Vercel, MongoDB, Sentry, and Cloudflare.

Fable's safety filters block benign math and nonsensical queries, prompting jokes from researchersThe blocked prompts cited cybersecurity and biology policies.

Jeff Bezos-backed industrial AI startup Prometheus raises $12 billion at a $41 billion valuation to automate physical-world engineeringThe startup aims to speed up hardware design by 10x.

OpenAI acquires Ona to run persistent, long-running AI agents in secure cloud environments through CodexCodex weekly active users grew 400% to 5 million.

Google releases DiffusionGemma, an open-source text diffusion model that generates over 1,000 tokens per second in parallel blocksQuantized GGUF versions run locally using llama.cpp

Pietro Schirano, Claude Engineer creator, uses Fable as an orchestrator and planner rather than for writing codeHe tasks other advanced models with writing the implementation.

Boris Power, OpenAI's head of applied research, takes a dual role as head of research for Thrive HoldingsThrive's 'Long Humans' thesis counters AI redundancy narratives.

ESM project founder Alex Rives and UC Berkeley collaborators launch a cryo-EM phase plate using the world's brightest continuous wave laserIt enables visualization of smaller proteins inside intact cells.

Altimeter Capital's Brad Gerstner and investor Gavin Baker analyze AI capital expenditure trends and a potential SpaceX IPOThey also evaluated gaming projects Fable 5 and Mythos.

Roboticist Marco Mascorro launches Use Computer, providing sub-second OS sandboxes on M4 Apple silicon to train computer-use agentsThe platform supports macOS, iOS, Windows, and Ubuntu.

Anthropic's Fable 5 achieves a 10x performance gain on Thoughtful Lab’s FrogsGame by autonomously training a weaker modelOther tested frontier models averaged under 4% on the benchmark.

Bearly AI creator Trung Phan's viral Knicks game meme parodies the cultural profiles of LinkedIn, X, Instagram, and RedditMichael Bloomberg represents LinkedIn while Kylie Jenner represents Instagram.

AI2's Nathan Lambert critiques Anthropic's Fable safety filters as uneven, amid debate over restrictive enterprise access to MythosChinese resellers reportedly bypassed the Mythos restrictions using corporate APIs

Peter Steinberger, OpenClaw steward, shares a workflow automating open-source repository maintenance via recurring five-minute AI agent loopsHe published the configurations in his agent-scripts repository.

Elon Musk predicts AI will render currency obsolete, necessitating a universal high incomeGary Marcus challenged the feasibility of Musk's redistribution claims.

Fable 5 Tops Benchmarks in Software Engineering, Research, and Vision