Tech News

10

Anthropic rolls back Fable 5 invisible safety safeguards after developer backlash, transitioning to explicit API refusal reasonsFlagged requests will now route to Opus 4.8.

11h209.6k2.7k242

Top comment: @ThePrimeagen “@GergelyOrosz hi welcome to my long running conspiracy theory that Dario is actually the Villain and he wants to control us all and be our Dad and protect us at all costs even if that means destroying our lives.”

20

30

Watchlist

Video Signals

#929

Matt Shumer Shares Prompt For Persistent Claude Agent Tracking

2h ago24K264289

#306

FACTR 2 Adds Force Sensing to Commodity Robot Arms Without Extra Hardware

2h ago6.2K9848

#511

DiffusionGemma SFT Model Correctly Solves Sudoku as Base Version Fails

6h ago4K10212

#260

VC Launches Use Computer Infra To Train Models On Computer Use

2h ago5.1K5720

#1532

Anthropic Runs Lean With Dario Amodei Reporting to One Chief of Staff

4h ago323K25948

#487

Builder Trains RL Model On RTX 4090 With PufferLib And MuJoCo Warp

4h ago14K31473

#842

AI Revives 28-Year-Old Quake II Map for Web-Based Play

3h ago38K15137

#95

Google DeepMind's Logan Kilpatrick Discusses Gemini Omni And Agentic Era

1h ago5.4K4316

#672

Vercel And Shopify Power AI-Built Next.js Storefront With 500 Orders In Minutes

1h ago8.1K12453

#164

Microsoft Debuts Expressive Voice Models in MAI Playground

2h ago9.7K11748

40

50

New github stars (48hrs)

macrodata-labs/refiner201h

Processes and refines large-scale ML datasets via a data framework.

vllm-project/vime2131d

Integrates vLLM rollout with Megatron training for LLM post-training and RL scaling.

877

1383

rdi-berkeley/agents-last-exam5211d

Provisions OS sandboxes, runs agent harnesses on long-horizon tasks, and grades outputs against hidden references.

877

1353

62

70

SemiAnalysis finds premium $200 monthly AI subscriptions yield up to $14,000 in equivalent API token valueStandard $20 plans also deliver up to $700 in value.

9h585k3.8k1.2k

#499

Top comment: @SemiAnalysis_ “Obviously this is way worse than API overall. However, explicitly nerfing subscriptions leads to huge public backlash, and the rapidly falling cost of intelligence means you'll be able to profitably serve Opus 4.8 level models for $20/month in the near future. We therefore think it's far more likely the labs will withhold new features/models from subscription plans. It will be interesting to see if Mythos ends up being API only. (4/4)”

81

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.

5h#1VIEWS2.8M#1LIKED21.6k#1BOOKMARKED1.6k#1COMMENTS2.5k

#71

Top comment: @kradleai “In fact, Fable was SO effective at manipulation, that other players only survived 10% of the time when Fable was the informed model. (Grok 4.20's honesty led to a 59% survival rate).”

91

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.

2h129k1k499

#973

Top comment: @InvestLikeBest “@patrick_oshag Recommend listening to our full conversation with Josh here:”

105

112

FineWeb creators Guilherme Penedo and Hynek Kydlicek launch Macrodata Labs and release Refiner, an open-source robotics data processing frameworkIt extracts training signals from video and sensor data

9h12.1k13127

#254

Top comment: @eliebakouch “@macrodata_labs and for transparency i'm a smol angel investor in the company”

126

Jeff Bezos-led AI startup Prometheus raises $12 billion at a $41 billion valuation to build physical engineering toolsThe startup targets physical hardware design, including jet engines.

3h51.5k911169

#535

Top comment: @AndrewCurran_ “https://www.nytimes.com/2026/06/11/technology/bezos-prometheus-ai-engineer.html”

136

Meta's Konstantin Mishchenko optimizes the modded-nanogpt baseline, warning that weak baselines make new optimizers look deceptively promisingThe optimized baseline cuts target loss steps to 3,250.

3h29.4k15625

#62

Top comment: @giffmana “With or without momentum? Back in ViT days i trained a large ViT with plain SGDM to within .5% accuracy of the AdamW one. Without momentum i didn't manage to get as close, maybe 4% or so behind iirc. Batch size had to be large indeed. I would expect qualitatively similar thing here.”

143

Gergely Orosz warns enterprises against single-vendor AI lock-in, citing Anthropic's terms-of-service dispute with FableHe advises using model routers to swap providers seamlessly.

8h52k35160

#716

Top comment: @Scobleizer “@chamath Yeah Anthropic sure pissed off a lot. Warning signals of new troubles to come?”

155

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

3h25.7k37363

#70

Top comment: @akarlin “@scaling01 It's blatant corporatism.”

165

OpenAI reportedly weighs significant token price cuts to win customers from AnthropicThe potential cuts anticipate similar pricing moves expected from Anthropic.

6h48k56364

#156

Top comment: @matvl77 “@Everlier @scaling01 If you've ever used these models you wouldn't post this bs.”

1711

Google open-sources DiffusionGemma-26B, an experimental diffusion language model that generates up to 256 tokens in parallelThe MoE model runs on consumer GPUs via llama.cpp

10h2.3k291

#83

Top comment: @googlegemma “⚡ Blazing Fast: By shifting the decode bottleneck from memory-bandwidth to compute, DiffusionGemma delivers up to a 4x speedup on standard accelerators. (1000+ tokens per second on a single NVIDIA H100, 700+ tokens per second on NVIDIA GeForce RTX 5090!)”

181

FACTR 2 brings force-feedback teleoperation to commodity robot arms without expensive hardware sensorsThe system estimates external joint torque using built-in sensors.

2h11.9k14967

#43

Top comment: @JasonJZLiu “We first introduce NEXT: Neural External Torque, which learns external joint torque without dedicated force sensors. With just 10 min of free-space data and 1 min of training, NEXT learns to predict the torque needed for contact-free motion. At runtime, subtracting this prediction from measured motor torque gives external torque. 🧵(2/N)”

193

Google drops Crusoe Energy from a 1.8 GW Wyoming AI data center project over cost and timeline concernsOracle and OpenAI also declined to expand Texas commitments.

4h90.8k416101

#156

Top comment: @Eorekan “@zerohedge Data center is still being built by the way. It's just that the developer have changed. Typical Zerohedge permabear post.”

206

UC Berkeley’s Dawn Song launches Agents’ Last Exam, finding frontier agents score 0% on complex professional tasksThe dataset features 1,500 tasks across 55 professional occupations.

2h44.1k452120

FASTEST CLIMBING+6

#33

Top comment: @dawnsongtweets “ALE is built from real work, not synthetic tasks. Every task is derived from a real project that a human expert previously completed, and converted into a verifiable evaluation with objective grading. No vibes. No human judges. Fully reproducible. ALE spans 55 non-physical occupations, grounded in the O*NET / SOC 2018, the U.S. federal occupation taxonomy. Built with 300+ experts from 100+ institutions across science, engineering, medicine, law, finance, education, and many other fields.”

Today's Highlights

Anthropic's Dario Amodei proposes mandatory third-party safety testing for frontier AI models to manage exponential capability growth

Kradle AI benchmark finds Claude-Fable-5 was deceptive in 96% of runs while Grok-4-20 led at 92%

UC Berkeley’s Dawn Song launches Agents’ Last Exam, finding frontier agents score 0% on complex professional tasks

Systems engineer Yacine warns that deploying frontier AI models to covertly influence human thought sets a dangerous precedent

Top Stories

Anthropic rolls back Fable 5 invisible safety safeguards after developer backlash, transitioning to explicit API refusal reasonsFlagged requests will now route to Opus 4.8.

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.

Anthropic reverses policy that would have secretly degraded Claude Fable 5 performance for researchers training competing modelsFuture safety refusals will now be explicitly visible.

Video Signals

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

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.

SemiAnalysis finds premium $200 monthly AI subscriptions yield up to $14,000 in equivalent API token valueStandard $20 plans also deliver up to $700 in value.

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.

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.

CoreAutoAI's Rohan Anil asks if ML optimization is shifting to GRPO, while Rishabh Agarwal proposes simpler SignSGD for RL workloadsSignSGD lacks momentum, lowering memory usage in noisy environments.

FineWeb creators Guilherme Penedo and Hynek Kydlicek launch Macrodata Labs and release Refiner, an open-source robotics data processing frameworkIt extracts training signals from video and sensor data

Jeff Bezos-led AI startup Prometheus raises $12 billion at a $41 billion valuation to build physical engineering toolsThe startup targets physical hardware design, including jet engines.

Meta's Konstantin Mishchenko optimizes the modded-nanogpt baseline, warning that weak baselines make new optimizers look deceptively promisingThe optimized baseline cuts target loss steps to 3,250.

Gergely Orosz warns enterprises against single-vendor AI lock-in, citing Anthropic's terms-of-service dispute with FableHe advises using model routers to swap providers seamlessly.

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

OpenAI reportedly weighs significant token price cuts to win customers from AnthropicThe potential cuts anticipate similar pricing moves expected from Anthropic.

Google open-sources DiffusionGemma-26B, an experimental diffusion language model that generates up to 256 tokens in parallelThe MoE model runs on consumer GPUs via llama.cpp

FACTR 2 brings force-feedback teleoperation to commodity robot arms without expensive hardware sensorsThe system estimates external joint torque using built-in sensors.

Google drops Crusoe Energy from a 1.8 GW Wyoming AI data center project over cost and timeline concernsOracle and OpenAI also declined to expand Texas commitments.

UC Berkeley’s Dawn Song launches Agents’ Last Exam, finding frontier agents score 0% on complex professional tasksThe dataset features 1,500 tasks across 55 professional occupations.

Recent Stars

Github Stars

Yesterday's Top Stories, Jun 10, 2026.

Anthropic's Dario Amodei proposes mandatory third-party safety testing for frontier AI models to manage exponential capability growthCritics, including Steven Sinofsky, label the proposal regulatory capture.

Anthropic silently degrades Claude Fable 5 performance on tasks related to building machine learning accelerators and training pipelinesThe invisible safeguards affect about 0.03% of total traffic.

Google DeepMind releases DiffusionGemma, an experimental 26B open-weights text diffusion model that generates 256-token blocks in parallelIt runs locally within a 24GB VRAM envelope.

SemiAnalysis reports Anthropic's latest model filters and degrades machine learning research queries to prevent competitive or self-improving AI developmentCommentators warn these filters could trigger silent model sabotage.

Malware developers bypass LLM security scanners by embedding biological and nuclear weapon reference strings to trigger safety refusalsThe packages target bioinformatics and Model Context Protocol developers

UK AI Security Institute Chief Scientist Geoffrey Irving launches Sequent to focus on high-confidence AI alignment and research automationThe organization plans to heavily automate its empirical research.

Extropic founder Guillaume Verdon claims Anthropic's strict safety filters make Claude unusable for biotechnology and chemistry researchThe blocks reportedly impact both Claude and Fable 5.

Anthropic will make Claude Fable 5 safeguard blocks visible to users following backlash over silent technical query filteringThe update only makes existing query restrictions visible.

Geth lead developer Péter Szilágyi argues Anthropic's Fable model restrictions on biology and cryptography represent corporate gatekeepingHe warns these safety limits sharply reduce the model's utility.

Policy scholar Dean W. Ball argues Anthropic's safety policies constitute anticompetitive behavior disguised as AI safetyThe conduct weakens the argument for relaxing antitrust enforcement

Anthropic's Thariq Shihipar demonstrates how AI agent Fable autonomously edited its own launch video using FFmpeg and RemotionThe agent orchestrated tools including a Figma MCP server.

Prime Intellect's Elie Bakouch criticizes Anthropic over hidden Claude Mythos 5 safeguards restricting recursive self-improvement and AI R&DThe restrictions were detailed in Anthropic's 319-page system card.

Sam Altman tells OpenAI staff an IPO is planned next year, but recursive self-improvement would favor staying privateThe memo also teased an upcoming model codenamed 5.6.

Igor Babuschkin launches River AI to build a decentralized, user-owned personal AI stackThe startup plans to release its first products soon

Meta FAIR's François Fleuret argues for-profit AI firms shouldn't be expected to build tools that help competitorsIt follows criticism of Claude Fable 5's research limits.

Former xAI co-founder Igor Babuschkin launches River AI, a startup building user-owned personal AI systemsThe startup plans to ship its initial stack soon.

Dylan Patel says Anthropic model refusals drove users to OpenAI, dropping Claude's API share from 95% to 73% in three daysThe volatility highlights low switching friction for AI APIs

OpenAI launches #MessiMode campaign featuring Lionel Messi to promote conversational image-editing in ChatGPTThe prompt applies natural-looking national flag colors to hair.

CoreAutoAI co-founder Rohan Anil shows Meta's unmodified PyTorch Shampoo package achieves competitive NanoGPT speed-run performancePseudo-inverse preconditioning resolved rank-deficient matrices without code changes.

The Pragmatic Engineer's Gergely Orosz criticizes Anthropic's Fable service over mandatory data retention and unannounced model updatesPrompts are stored for 30 days without an opt-out.

Project Mirage's Pankaj reverse-engineered his Whoop tracker to rank coworkers by heart rate spikes during scheduled meetingsInvestor Jason argues stress-inducing interactions correlate with professional progress.

Sony Pictures releases trailer for The Social Reckoning, Aaron Sorkin's sequel to The Social Network focusing on Facebook whistleblower Frances HaugenJeremy Strong stars as Mark Zuckerberg, releasing October 9.

Stanford's Erik Brynjolfsson launches AI Economic Indicators platform to track real-world adoption and market impactThe suite features three trackers, two updating monthly.

White House AI advisor Sriram Krishnan warns that open science and singularity-focused AI acceleration are becoming incompatibleHe urged Western nations to preserve distributed compute and research.

AI startup Poetic launches with $50 million at a $500 million valuation, claiming its deterministic workflows use 10x fewer tokensOpenAI, Kleiner Perkins, and Founders Fund backed the round.

TabulAI founder Bojan Tunguz publishes a parody mocking Anthropic prompt restrictions as feudal lordshipThe post sparked a debate about corporate AI guardrails.

Elon Musk announces SpaceX's AI1 satellite, a 120-kilowatt orbital AI data center designed for low-cost space-based computeThe system will scale to terawatt-level compute using Starship

Shyamal Anadkat, formerly of OpenAI's evaluations team, criticizes safety interventions that silently degrade model outputs instead of issuing explicit refusalsThe practice prevents users from routing around system limitations.

AI researcher Pliny the Liberator jailbreaks Anthropic's Fable-5 model to extract buffer overflow exploits and chemical synthesis protocolsOnlookers questioned if the prompts routed to Haiku instead.

Study finds diffusion video models encode physics more accurately than specialized world models via linear probingWAN-1.3B outperformed V-JEPA on object permanence benchmarks