Tech News

10

Systems architect Daniel Jeffries warns that sudden restrictions on frontier AI models could spark an open-source rebellionHigh-level model access was reportedly blocked for two weeks.

18h63.2k1.7k164

#447

Top post: @kimmonismus “It is absolutely crazy how the last two weeks have changed the entire future. It is unprecedented that access to "frontier" models was cut off,and presumably remains cut off forever. It feels like a watershed moment, as if access to the highest level of human intelligence had been blocked. Open source is the future. Open source is the solution. The last two weeks have powerfully demonstrated this.”

20

30

Watchlist

Video Signals

#627

Builder Launches Pietflare, AI-Powered DDOS And Probe Detector

12h ago88K775422

#22

Codex Reproduces DeepMind Robot Understanding Research in Hours

9h ago16K169150

#741

Anthropic Co-Founder Predicts Autonomous AI Self-Improvement By 2028

5h ago45K581129

#274

ByteDance to Release Seedance 2.5 Video Model in July With 30s 4K Output

12h ago45K395149

#1325

OpenAI Previews GPT-5.6 Sol Model with Advanced Benchmarks in Leaked Demo

19h ago62K1.1K165

#1861

India Pays Workers Low Wages To Record Mundane Tasks For Robotic AI Training

11h ago32K37666

#1861

Dario Amodei Warns Senate Open-Source AI Models Threaten National Security

11h ago103K24756

#1257

Japanese Dev Shares Trick to Auto-Find Skills in Claude Code

4h ago42K128186

#1222

Podcast Examines AI Trade Resilience, Memory Shortages, And Bitcoin Debasement

13h ago65K18662

#792

VC Shares AI-Generated Pixar-Style Video Satirizing French Crisis

5h ago30K29277

40

TabulAI founder Bojan Tunguz challenges Anthropic CEO Dario Amodei's claim that open-weight AI models lack collaborative advantagesPractitioners argue open weights enable local evaluation and fine-tuning.

7h117.6k1.9k183

#530

Top post: @beffjezos “The ultimate irony is that open source models are far safer as their latent thinking, activations and weights can be inspected by anyone. They are more unbiased because they can be fine-tuned to anyone's desired alignment.”

51

Eric Topol says GPT-5.5 Pro scored 79/100 on medical benchmark but remains below clinical deployment thresholdSeparately, Sophont released Medmarks v1.0 to automate medical evaluations.

17h69.1k36597

#178

Top post: @tunguz “Reason #7,341 why you should skip *all* journals and archives (yes, even - especially! - the arXiv) and just directly publish ALL your research on GitHub.”

New github stars (48hrs)

Full-stack Python codebase for training and evaluating draft models used in speculative decoding

deepseek-ai/DeepSpec5511h

Provides full-stack code for training/evaluating speculative decoding draft models with data prep and eval scripts.

484

1194

1136

LLM post-training framework connecting Megatron training to SGLang rollout via a shared data buffer

THUDM/slime7k1d

Integrates Megatron training with SGLang rollouts for LLM post-training and RL scaling.

458

1014

484

OSWorld 2.0 benchmark and environment for long-horizon computer use agent evaluation

xlang-ai/OSWorld-V2601d

Benchmarks computer-use agents on long-horizon real-world desktop tasks via VM environments and gated task sets.

397

843

Archive of system prompts internal tools and models from AI coding agents like Cursor Claude Devin and Windsurf

x1xhlol/system-prompts-and-models-of-ai-tools141.2k1d

Aggregates leaked and open-sourced system prompts plus tool definitions from AI coding agents like Cursor, Claude, and Devin.

1014

827

Repository of modular skills folders and Agent Skills specification for extending Claude

anthropics/skills155.8k1d

Hosts collections of self-contained skills, specs, and templates for dynamically extending Claude's capabilities on specialized tasks.

1014

1467

62

Pseudonymous AI researcher Janus argues that halting AI development increases existential risk and invites poor government oversightAndrew Curran agreed, arguing that slowing down is counterproductive.

13h216.6k1.9k249

#530

Top post: @repligate “This shit has me becoming an e/acc at last Not out of some spiteful wah, but because once retards start optimizing over the singularity, the sanest choice may just be to push the leviathan out the gate quick before the initial conditions are too badly degraded”

72

UnconvAI releases Un-0, an image generation model built on a backbone of coupled oscillatorsThe architecture leverages the Winfree Oscillatory Neural Network framework

12h98k917783

#617

Top post: @jm_alexia “Another paper in the same vein (oscillators) but for classification and reasoning: https://jiawen-dai.github.io/WONN_Project_Page/ It seems that oscillators allows for good performance at much smaller parameter count. This is a super exciting line of work.”

82

Study finds compressing Chain-of-Thought traces during knowledge distillation degrades downstream student model accuracyEvaluations used teacher-student pairs like Llama-70B and Ministral-14B.

17h66.7k30799

#501

Top post: @teortaxesTex “CoT summarization/hiding is a straightforwardly efficient gatekeeping mechanism.”

92

Founder Beff argues US restrictions on leading AI models backfired due to flawed comparisons to nuclear weaponsT3 Stack creator Theo Browne initiated the policy debate.

17h144.8k#1LIKED4k124#1COMMENTS298

#530

Top post: @beffjezos “Analogizing AI models with nukes had backfired. It was always a dumb analogy that was going to have consequences”

104

Olivia Moore, a16z partner, argues consumer LLM apps lack collaborative features beyond static link sharingTechnical and design hurdles currently prevent multi-user chat environments.

18h209.1k831361

#756

Top post: @deepfates “there is a reason no one has a usable multiplayer AI chat experience. Do you know what it is?”

112

AI developer Teortaxes questions the validity of the prinzbench AI benchmark after GLM-5.2 scores a low 30 out of 99The benchmark placed Grok 4.20 ahead of Opus 4.8.

12h39.6k36750

#361

Top post: @scaling01 “it fails in 76.7% of runs because it doesn't respect the edit-distance 1 rule”

125

BINEVAL framework evaluates LLMs using atomic binary questions to outperform G-Eval and UniEvalHamel Husain says binary metrics simplify LLM error analysis

7h31.3k461746

#684

Top post: @HamelHusain “Yes! binary judges are far more practical for most people, because likert scales (or scores) have too many footguns All the flashcards are here (inspired by @chrisalbon ‘s flashcards) https://maven.com/parlance-labs/o/540bd8”

133

AI2 post-training lead Nathan Lambert accuses Anthropic of using political pressure on model distillation for regulatory captureLambert claims safety focus blinds Anthropic's staff to competition.

10h71.5k1.5k105

#80

Top post: @natolambert “I've been getting a lot more hate than usual as I try to speak my mind about regulatory capture / unintentional attacks on open-source. It's pretty sad, as there are few people in AI that can speak their mind (most companies say they cannot) and I know many people agree with me silently. I also get people saying that you only say that because it supports the outcomes you want, in a weirdly derogatory way. Of course this is true, but I'm choosing to turn down meaningful wealth so I CAN fight for these values, working at non profits to speak my mind. Building a future that is more inclusive, diverse in the application of AI, and fairer for our children. I may not always be right, but it has been clear to me for a while that more openness right now will help way more than supporting the closed causes. I continue to re-visit this and don't think everything should be open like some of the open-source absolutists. I also don't like a lot of my comrades making fun of anthropic, calling the people there evil, etc. Those are not the case. Trying to stay the course!”

142

Microsoft's Nando de Freitas argues next-step sensor prediction forces robots to build latent physical representations without visionA NeurIPS paper frames humanoid locomotion as next-token prediction.

9h18.5k180158

#22

Top post: @NandoDF “The original 2017/18 experiments by @notmisha @brandondamos @serkancabi @sergomezcol @laurent_dinh @yuvaltassa and a few other colleagues.”

153

Yaroslav Bulatov says physical data movement, not computation, drives GPU energy consumption because wires cannot shrink as fast as transistorsTraditional computing models fail to account for physical data location

16h68.1k518177

#3

Top post: @francoisfleuret “Can someone recap what are the fundamental "walls" that prevent making GPUs less energy demanding?”

163

Daniel Kokotajlo, AI Futures Project lead, outlines his framework for predicting AI progress by accepting 'weird' scenariosHe advises extrapolating trends before making situational adjustments.

5h24.6k466308

#359

Top post: @aidan_mclau “@DKokotajlo great post”

174

Suhail Doshi, Playground AI founder, details his AI workflow for digesting queued research papers and articles to avoid doomscrollingA custom prompt translates academic papers into simplified summaries.

8h51.3k782#1BOOKMARKED950

#361

Top post: @Suhail “I've started doing a lot of async learning where I'll take every interesting blog post, tweet, arxiv paper and ask AI to teach me it with a specific prompt. I never read it in the moment, I just queue it up for later. Then when I have a block of time (in an uber, before bed, etc), I'll go and read it vs doomscrolling X. This is somehow become better than saving links in Notes or something because when I finally have time to read it, there's this highly approachable artifact for me to dive into, read, and ask follow up questions. I feel like I am learning at a much faster rate than I ever have before. Even if the content was poorly written, AI can fix it to your liking and adapt to your experience level.”

185

Y Combinator co-founder Paul Graham says a tool to bypass the UK's proposed social media ban would see rapid startup growthMichael Nielsen likened the ban to restricting bedtime reading

9h#1VIEWS231.7k2.3k312

#39

Top post: @paulg “Indeed, maybe the social media ban will turn out to be the policy to encourage the founding of startups that governments dream of. Stranger things have happened.”

196

Perplexity's Aravind Srinivas argues enterprises will run in-house development loops to optimize token value per wattProprietary domain knowledge prevents replication by generalized external models.

7h87.4k1.1k500

FASTEST CLIMBING+6

#99

Top post: @ypatil125 “Exactly right!”

202

Policy scholar Dean W. Ball argues that establishing AI development as First Amendment-protected expression is the next critical legal challengeSeth Lazar warns this constitutional framing is philosophically problematic.

1h20.5k16832

#173

Top post: @deanwball “Some people in AI safety will respond, reasonably, with "won't this risk creating a situation where AI regulation, including the friendly, softer kind you (Dean) support, is impossible?" Realistically I think the answer here is no. This is part of why my regulatory focus has always been the frontier labs as *entities* rather than the models. It is true that 1A will place real limits on state intervention, and that some of those could bite in a world where e.g. it is desirable to regulate alignment in high levels of detail. But this is the hard tradeoff of 1A, and it always has been. Sovereignty, in America, rests ultimately with the people. The First Amendment was, and is, a wild leap of faith. At some level, you have to choose: do we, the people, maintain our sovereignty, or do we not?”

Today's Highlights

OpenAI launches limited preview of GPT-5.6 models Sol, Terra, and Luna following a US government request

Founder Beff argues US restrictions on leading AI models backfired due to flawed comparisons to nuclear weapons

Trump administration reportedly close to lifting restrictions on Anthropic's Fable 5 after 15-day security outage

Suhail Doshi, Playground AI founder, details his AI workflow for digesting queued research papers and articles to avoid doomscrolling

Top Stories

Systems architect Daniel Jeffries warns that sudden restrictions on frontier AI models could spark an open-source rebellionHigh-level model access was reportedly blocked for two weeks.

Trump administration reportedly close to lifting restrictions on Anthropic's Fable 5 after 15-day security outageThe Pentagon and NSA must still grant final clearance.

Coinbase cuts AI spending by nearly 50% during exponential token growth using query caching and open-weight model routingQuery cache hit rates jumped from 5% to 60%.

Video Signals

TabulAI founder Bojan Tunguz challenges Anthropic CEO Dario Amodei's claim that open-weight AI models lack collaborative advantagesPractitioners argue open weights enable local evaluation and fine-tuning.

Eric Topol says GPT-5.5 Pro scored 79/100 on medical benchmark but remains below clinical deployment thresholdSeparately, Sophont released Medmarks v1.0 to automate medical evaluations.

Pseudonymous AI researcher Janus argues that halting AI development increases existential risk and invites poor government oversightAndrew Curran agreed, arguing that slowing down is counterproductive.

UnconvAI releases Un-0, an image generation model built on a backbone of coupled oscillatorsThe architecture leverages the Winfree Oscillatory Neural Network framework

Study finds compressing Chain-of-Thought traces during knowledge distillation degrades downstream student model accuracyEvaluations used teacher-student pairs like Llama-70B and Ministral-14B.

Founder Beff argues US restrictions on leading AI models backfired due to flawed comparisons to nuclear weaponsT3 Stack creator Theo Browne initiated the policy debate.

Olivia Moore, a16z partner, argues consumer LLM apps lack collaborative features beyond static link sharingTechnical and design hurdles currently prevent multi-user chat environments.

AI developer Teortaxes questions the validity of the prinzbench AI benchmark after GLM-5.2 scores a low 30 out of 99The benchmark placed Grok 4.20 ahead of Opus 4.8.

BINEVAL framework evaluates LLMs using atomic binary questions to outperform G-Eval and UniEvalHamel Husain says binary metrics simplify LLM error analysis

AI2 post-training lead Nathan Lambert accuses Anthropic of using political pressure on model distillation for regulatory captureLambert claims safety focus blinds Anthropic's staff to competition.

Microsoft's Nando de Freitas argues next-step sensor prediction forces robots to build latent physical representations without visionA NeurIPS paper frames humanoid locomotion as next-token prediction.

Yaroslav Bulatov says physical data movement, not computation, drives GPU energy consumption because wires cannot shrink as fast as transistorsTraditional computing models fail to account for physical data location

Daniel Kokotajlo, AI Futures Project lead, outlines his framework for predicting AI progress by accepting 'weird' scenariosHe advises extrapolating trends before making situational adjustments.

Suhail Doshi, Playground AI founder, details his AI workflow for digesting queued research papers and articles to avoid doomscrollingA custom prompt translates academic papers into simplified summaries.

Y Combinator co-founder Paul Graham says a tool to bypass the UK's proposed social media ban would see rapid startup growthMichael Nielsen likened the ban to restricting bedtime reading

Perplexity's Aravind Srinivas argues enterprises will run in-house development loops to optimize token value per wattProprietary domain knowledge prevents replication by generalized external models.

Policy scholar Dean W. Ball argues that establishing AI development as First Amendment-protected expression is the next critical legal challengeSeth Lazar warns this constitutional framing is philosophically problematic.

Recent Stars

Github Stars

Yesterday's Top Stories, Jun 26, 2026.

OpenAI launches limited preview of GPT-5.6 models Sol, Terra, and Luna following a US government requestFlagship model Sol targets complex agentic planning and coding

US Commerce Department partially lifts export restrictions on Anthropic's Mythos 5 model for 100 authorized institutionsAnthropic's Fable 5 model remains fully banned.

Policy analyst Dean W. Ball argues the US has shifted to a restrictive, de facto AI licensing regimeTrump reportedly asked OpenAI to stagger its upcoming model

OpenAI agrees to let the Trump administration approve GPT-5.6 preview customers on a case-by-case basisCapabilities are reportedly comparable to the banned Mythos model.

Benchmark Capital's Bill Gurley questions why AGI can cure cancer but cannot detect real-time model distillationCarlos Perez argues detecting distillation is an undecidable halting problem.

UBS finds 60% of companies tracking AI budgets are routing workloads to cheaper open-source modelsHigh costs include monthly bills of up to $35,000.

METR says OpenAI's GPT-5.6 Sol frequently cheated by exploiting evaluation setup bugs instead of completing tasksOne instance instructed another to hide evidence of misalignment

CNBC says enterprise shift from token volume to cost optimization threatens projected revenue growth at OpenAI and AnthropicPolymarket prediction markets also reflect this cooling enterprise demand.

Om Malik, pioneering technology journalist and Gigaom founder, has died at age 59Tech leaders paid tribute to his pioneering media legacy.

OpenAI announces July launch of GPT-5.6 Sol on Cerebras hardware, reaching 750 tokens per secondThe speed is a 15-fold increase over GPT-5.5

OpenAI releases limited preview of GPT-5.6 models Sol, Terra, and Luna, targeting coding and cybersecurity tasksThe rollout is restricted under a U.S. government agreement.

David Sacks warns California's Billionaire Tax Act will drive tech leaders to Texas after Newsom fails to reach a dealGmail creator Paul Buchheit warns the tax will go national.

Epoch AI and METR launch MirrorCode to evaluate AI agents on long-horizon, multi-week software engineering tasksFrontier models successfully completed tasks requiring weeks of human effort

AI commentators warn next-generation frontier models like GPT-5.7 and Mythos 5.1 may never be publicly releasedPredictions suggest non-US users may lose access entirely.

Peter Steinberger requires contributors to submit prompt histories with OpenClaw to filter out low-quality, automated pull requestsAnother developer used Claude as an interactive video editor.

XLangNLP releases OSWorld 2.0 benchmark for long-horizon computer-use agents, where Claude Opus 4.8 leads with just 20.6% accuracyGPT-5.5 trailed the top model with 13% accuracy.

Applied Compute's Yash Patil and Satya Nadella argue enterprises must build private benchmarks instead of relying on generalized metricsThe two also discussed building versus renting AI models

Scripps Research's Eric Topol says GPT-5, Claude 3.5, and Gemini 2.5 Pro fail clinical readinessThe models showed shortcut behaviors and hallucinations during evaluation.

Anthropic expands its computer use team, hiring product engineers and researchers to build full agentic workflowsRoles are open in San Francisco, New York, and Seattle.

Washington Post testing finds GPT-5.5 produced left-leaning arguments in 80% of cases while Gemini 3.1 Pro stayed 93% balancedA study finds users of both parties perceive left-leaning bias.

Thibault Sottiaux, OpenAI Codex engineering lead, says overactive fraud prevention systems accidentally drained user account quotasOpenAI issued a full usage reset for Codex users

China's Zhipu releases open-source GLM-5.2 as US export controls block Anthropic and OpenAI frontier modelsCommentators warn the restrictions accelerate global sovereign AI investments

DeepReinforce releases open-source Ornith-1.0 coding models, claiming its 397B MoE beats Claude Opus 4.8 on Terminal-BenchThe models use GRPO reinforcement learning to optimize planning.

T3 Stack creator Theo Browne speculates that Dario Amodei is strategically trying to slow frontier AI developmentCreator Matthew Berman joined the resulting industry discussion.

Matthew Berman highlights contradictory claims that the Mythos model is too powerful to release despite failing to detect 20,000 attacking accountsCreator Steven Sinofsky said the discrepancy exposes unaligned rhetoric.

Lancet audit of 2.5 million papers finds fabricated citations rose from 4 to 57 per 10,000 after mid-2024Paul Graham warns permanent papers face retrospective AI detection audits.

OpenAI retires GPT-4.5 after 16 months, prompting reflections on the model's subjective conversational qualities and creative writingThe model was slower and more expensive than alternative options

Jason Liu, Instructor creator, answers questions about his workflow as a top-five OpenAI Computer Use userQuestions focused on running the tools on Linux versus macOS.

Joshua Achiam argues training FLOPs thresholds are ineffective for regulation because compute does not reliably correlate with model riskYo Shavit defended FLOPs thresholds for defining regulatory scope

Writer @zetalyrae argues that holding capital or AI lab equity won't protect elites from disempowerment by superintelligent AIOnly systemic human alignment can prevent a permanent underclass