Tech News

10

OpenAI Publishes GPT-5.6 Preview System Card

28m#1VIEWS188.4k#1LIKED3.1k169#1COMMENTS444

Top comment: @OpenAI “We believe in broad access and plan to make GPT-5.6 Sol, Terra, and Luna generally available in the coming weeks. For now, at the request of the U.S. government, we’re starting with a limited preview among a small group of trusted partners in Codex and the API.”

20

30

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.

4h161k2.5k266

#178

Top post: @packyM “a machine that was really of loving grace would stop the chinese from distilling it illicitly”

Watchlist

Video Signals

#219

Yash Patil Urges One AI Model Per Company In Nadella Interview

1h ago19K190103

#530

Sam Altman Blames Dario Amodei for Spooking Regulators and Slowing Frontier AI

9h ago15K25431

#831

Chart Compares AI Data Centers Water Use to Global Agriculture and Industry

7h ago84K00

#60

Podcast Examines Next AI Training Paradigm and Lab Research Bets

43m ago7.8K7777

#350

New England Tech Scene Unveils AI Brain Platforms, Robot Submarines And Cancer Cures

2h ago4.8K4521

#33

DanceOPD Introduces On-Policy Generative Field Distillation for T2I Models

2h ago4.9K3215

#168

Reid Hoffman Says AI Offers Best Chance to Cure Cancer

1h ago6.3K256

#1853

Omni AI Generates Realistic Video From Image References

6h ago9.1K7824

#1205

Alex Mordvintsev Teases Upcoming Release with Evolving Dot Animations

1h ago2.3K6310

#543

Anthropic Engineers Ship 8x More Code With AI Coding Tools

1h ago5.5K1520

40

Epoch AI releases MirrorCode to test AI systems on multi-day software engineering tasksThe framework helps researchers study model inference compute efficiency.

2h25.6k32387

#184

Top comment: @emollick “Paper: https://epoch.ai/files/MirrorCode_3d9aab.pdf”

51

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.

8h102.4k823#1BOOKMARKED362

#816

Top post: @EricTopol “Link for free access https://rdcu.be/fqznS This extensive assessment work was led by @hoifungpoon and Yu (Aiden) Gu”

New github stars (48hrs)

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

THUDM/slime6.8k24m

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

1014

484

90

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

xlang-ai/OSWorld-V21627m

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-tools141.2k9h

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/skills155.1k9h

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

1014

1467

Web-based benchmark platform evaluating MLLM agents through diverse dynamic CAPTCHA puzzles

MetaAgentX/OpenCaptchaWorld789h

Hosts a Flask-based web platform with 20 CAPTCHA types to benchmark MLLM agents on visual reasoning and interaction tasks.

1014

939

61

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.

4h98.4k1.2k123

#39

Top post: @paulg “Academia is uniquely vulnerable to being caught out in this way, because the test of productivity is publication. So all the evidence is out there, permanently, to be analyzed by ever more sophisticated tools.”

72

DeepSeek raises $7.4 billion at a $50 billion valuation to compete with Anthropic's Mythos modelThe startup plans to double its 300-person workforce.

4h123.7k1.1k132

#1257

Top post: @kimmonismus “I think many people are not yet aware of the tectonic shift taking place. By preventing state-of-the-art capabilities - at least insofar as we are able to use them - open source becomes not only more attractive for one’s own applications, but more attractive overall. This also applies, for example, to entire states such as the European Union, provided it genuinely sets itself the goal of implementing AI. This not only attracts new investments and fresh capital, but also creates an opportunity for a PR coup that draws many idealists toward open-source companies. In that sense, I can well imagine that companies like OpenAI and Anthropic will suffer the most from regulation by the U.S. government - on two fronts - and that this indirectly does open source a tremendous service, especially given that it now primarily comes from China.”

83

XLANG Lab releases OSWorld 2.0, a computer-use agent benchmark where frontier models achieve just 20.6% accuracyWorkflows require an average of 318 tool calls each.

2h16.1k14327

#326

Top post: @JustinLin610 “i still remember the discussion of the case of us visa application and i thought that must be mission impossible for ai and just said u guys must be crazy... now it seems that i am the dumb hahah! anyway, time for frontier models to fight again! bon courage!”

94

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

59m19.8k198103

#219

Top post: @lulumeservey “@ypatil125 The misty scenery outside looks like Pandora from the Avatar movies!”

106

A generative diffusion model trained on recipes reconstructs the Big Mac and designs optimized, sustainable burgersThe model learned recipe relationships without explicit supervision.

5h50k24690

#184

Top post: @emollick “Finally, AI finds its ultimate uncontroversial use. A diffusion model trained on burger recipes "discovers the classic Big Mac without explicit supervision and generates novel burgers optimized for deliciousness, sustainability, or nutrition." ASI= automated slider intelligence”

114

Developer @xlr8harder proposes testing recursive loops after Digg scrapes and AI-summarizes niche Twitter postsThe experiment measures data degradation from recursive AI loops

10h16.5k1555

#1014

Top comment: @xlr8harder “@deepfates open source model fine tuner xlr8harder indignantly shares digg update with creator deepfates, in which digg's summary bot slanders their recent posts as jokes”

125

Investigators blame Russian state actors for Jaguar Land Rover cyberattack that caused $2.5 billion in damagesThe incident halted UK vehicle production for a month

4h71.9k51229

#797

Top post: @bscholl “According to my sources, the Jaguar hackers planted a virus in the CAD system that deleted their real designs and left behind something hideous”

137

Zvi Mowshowitz argues the new White House AI policy allows ad hoc control over frontier AI model developmentCritics blame the ad hoc policy on tech-led deregulation

2h15.9k46119

RISING LIKES461

#1396

Top post: @binarybits “This seems like a weird way for Marc Andreessen to acknowledge that supporting Trump was a mistake but I'll take it!”

1413

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.

1h6.8k12588

FASTEST CLIMBING+13

#1262

Top post: @ehsanik “Here is the job description and link to apply for those interested: https://job-boards.greenhouse.io/anthropic/jobs/5238637008”

159

OpenAI plans a July launch for GPT-5.6 Sol on Cerebras hardware at 750 tokens per secondGPT-5.6 Sol Ultra scored 91.9% on TerminalBench 2.1.

16m10.3k13214

#586

Top post: @Yuchenj_UW “GPT-5.6 is finally coming. GPT-5.6 Sol beats Claude Mythos 5 on TerminalBench. And on Cerebras, GPT-5.6 Sol can reach up to 750 tokens per second. Pretty fast for a model of this size. Now I just hope it can be rolled out to everyone.”

1613

Satirist Attempts Joke Prompt Injection On Digg AI Scraper

10h11k905

#1014

Top comment: @granawkins “@deepfates incredible, this is the best art i've seen since that woman hung upside down naked inside a giant bell and bonged herself back and forth with her tits hanging out the bottom of it”

175

Google DeepMind's Susan Zhang and Alexander Doria flag benchmark discrepancies in Mistral's OCR 4 launch comparisonsThe charts underreported Chandra 2 and omitted Infinity Parser.

5h12.4k10115

#84

Top post: @deliprao “Vik has been doing fantastic work on PDFs for years and Datalab's suite of tools (including Marker and Chandra) carried us through multiple DARPA evaluations. Weird to see them misrepresented, but I cannot recommend Chandra highly enough if you are dealing with complex PDFs.”

188

Donald Trump announces a 100% tariff on any country that imposes a Digital Services Tax on US tech platformsThe measure targets European nations discussing new digital taxes

51m18.4k15514

#178

Top post: @GaryMarcus “Always playing checkers and never chess.”

196

GPU provider Iren's reported $50 million Warriors jersey sponsorship sparks debate over B2B marketing ROIStadium suite perks could close major enterprise GPU contracts.

3h31.2k16510

#1157

Top post: @TurnerNovak “Don’t you get other benefits? Like premium suite at the games? (Putting aside if the company/product is legit or not), all they’d need is it to close a single compute deal with someone for it to pay off. Feels possible if they’re inviting 5-8 execs from every tech co to every home game.”

208

OpenAI's technical lead for compute returns from Meta, sparking debate over rapid AI skill depreciationCommentators cited Meta's challenging internal culture for the departure.

9h15.2k10318

#447

Top post: @Scobleizer “It is a very difficult culture at Meta to navigate at the moment, I have heard from many a current and former employee. Plus Apple has Formula 1. Which means all the rich people aren’t moving ecosystems easily. And why every Apple exec I have talked to is not worried. Knows it can come in very late and still take the market. OpenAI has a shot at disrupting Apple. Meta? Not so much. Audio headsets are only a temporary game. But they could last for years. The real game is AI that can see and show. Requires glasses with color displays in both eyes. The video I just did with a Chinese AI glasses startup is very revealing as to marketing as well as the digital eyewear market.”

Today's Highlights

Trump administration asks OpenAI to stagger the rollout of GPT-5.6 and restrict access to approved customers

OpenAI Publishes GPT-5.6 Preview System Card

Scripps Research's Eric Topol says GPT-5, Claude 3.5, and Gemini 2.5 Pro fail clinical readiness

Zvi Mowshowitz argues the new White House AI policy allows ad hoc control over frontier AI model development

Top Stories

OpenAI Publishes GPT-5.6 Preview System Card

Dean W. Ball warns US AI policy has rapidly become highly restrictive, proposing 35 reforms to replace improvised model licensingAI safety expert Yo Shavit endorsed Ball's policy solutions.

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.

Video Signals

Epoch AI releases MirrorCode to test AI systems on multi-day software engineering tasksThe framework helps researchers study model inference compute efficiency.

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.

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.

DeepSeek raises $7.4 billion at a $50 billion valuation to compete with Anthropic's Mythos modelThe startup plans to double its 300-person workforce.

XLANG Lab releases OSWorld 2.0, a computer-use agent benchmark where frontier models achieve just 20.6% accuracyWorkflows require an average of 318 tool calls each.

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

A generative diffusion model trained on recipes reconstructs the Big Mac and designs optimized, sustainable burgersThe model learned recipe relationships without explicit supervision.

Developer @xlr8harder proposes testing recursive loops after Digg scrapes and AI-summarizes niche Twitter postsThe experiment measures data degradation from recursive AI loops

Investigators blame Russian state actors for Jaguar Land Rover cyberattack that caused $2.5 billion in damagesThe incident halted UK vehicle production for a month

Zvi Mowshowitz argues the new White House AI policy allows ad hoc control over frontier AI model developmentCritics blame the ad hoc policy on tech-led deregulation

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.

OpenAI plans a July launch for GPT-5.6 Sol on Cerebras hardware at 750 tokens per secondGPT-5.6 Sol Ultra scored 91.9% on TerminalBench 2.1.

Satirist Attempts Joke Prompt Injection On Digg AI Scraper

Google DeepMind's Susan Zhang and Alexander Doria flag benchmark discrepancies in Mistral's OCR 4 launch comparisonsThe charts underreported Chandra 2 and omitted Infinity Parser.

Donald Trump announces a 100% tariff on any country that imposes a Digital Services Tax on US tech platformsThe measure targets European nations discussing new digital taxes

GPU provider Iren's reported $50 million Warriors jersey sponsorship sparks debate over B2B marketing ROIStadium suite perks could close major enterprise GPU contracts.

OpenAI's technical lead for compute returns from Meta, sparking debate over rapid AI skill depreciationCommentators cited Meta's challenging internal culture for the departure.

Recent Stars

Github Stars

Yesterday's Top Stories, Jun 25, 2026.

Trump administration asks OpenAI to stagger the rollout of GPT-5.6 and restrict access to approved customersOpenAI agreed to the temporary case-by-case preview structure

Om Malik, GigaOM Founder and Tech Journalist, Passes Away

OpenAI research shows 98% of its employees use Codex agents, with over 25% of tasks exceeding eight hoursMedian output tokens for the research team increased 56-fold.

Apple raises Mac and iPad base prices by up to 25% due to rising memory chip costsThe price hikes average nearly 20% across consumer devices.

Hugging Face CEO Clément Delangue says the open AI platform crossed $100 million in annual run-rateThe platform remains free for 97% of its users.

Anthropic says Alibaba used 25,000 fraudulent accounts to run a massive distillation attack on ClaudeThe operators targeted Claude's software engineering and agentic capabilities.

Paul Graham argues writing without AI is like running for exercise, rejecting claims that avoiding LLMs is obsoleteResearcher Michael Nielsen warns AI adoption could diminish human thinking.

Exponential View report finds the generative AI economy generated $110 billion in sales over the past yearAnnualized revenue run rate now exceeds $175 billion

Elon Musk calls for a complete redesign of the X recommendation algorithm, sparking debate over prioritizing network connectionsGauntlet AI's Austen Allred requested more followed-account visibility.

Rohit Krishnan warns unauthorized Chinese resellers offering heavily discounted Claude API access may be committing model-substitution fraudResellers allegedly swap Claude Opus with GLM or DeepSeek.

Aleph Neuro releases highest-resolution noninvasive 3D human brain scans captured using ultrasound and microbubble injectionsThe data will help develop noninvasive brain-computer interfaces.

Virtue AI founder Dawn Song joins Meta Superintelligence Labs as VP of AI Research alongside her startup's teamVirtue AI developed security benchmarks for trustworthy agentic systems.

Mark Cuban argues data center opposition is a proxy for AI resentment, advising LLM firms to directly fund local transition programsSteven Sinofsky agreed, calling the backlash generalized tech resentment.

Thinking Machines Lab co-founder Lilian Weng reconciles Kaplan and Chinchilla scaling laws for optimizing AI compute allocationThe framework maps training loss along a power-law curve.

X begins rolling out X Money neobanking services and metal Visa debit cards to US Premium+ subscribersThe service provides up to $10 million FDIC insurance

Databricks claims top GLM-5.2 inference speed on the Artificial Analysis leaderboard, hitting 392 tokens per second to beat Fireworks AIDeveloper Taelin questioned whether the benchmark used model quantization

IBM announces 0.7nm process node using 3D NanoStack architecture, marking the first sub-1nm chip technologyIt cuts energy consumption by 70% compared to 2nm.

Dell shareholders vote with 97% approval to relocate the company's legal domicile from Delaware to TexasCEO Michael Dell cited Texas's talent and business climate

Naveen Rao's Unconventional AI releases Un-0, an image model using coupled oscillators instead of neural networksThe physics-based approach generates 64x64 ImageNet samples.

Ramp names co-founders co-CEOs and promotes former applied AI head Rahul Sengottuvelu to CTOHamid Dadkhah was also named head of engineering.

Sazabi raises $8 million to build a self-healing observability platform for code generated by AI agentsThe system uses logs to autonomously resolve production issues.

Rippling CEO Parker Conrad launches Rippling Data Cloud to ground natural-language queries and reduce token costsIt queries organizational and personnel structures using plain language.

Paul Graham argues a tech company's shipping rate predicts success, while investor Alfred Lin says speed needs directionGraham claims this execution metric holds from startups to giants

Google restructures its AI coding strike team into a "midtraining" group to boost Gemini's performanceThe group focuses on the phase between pretraining and post-training

xAI founder Elon Musk jokes about naming a potential AI regulatory body the "AI Associated Institute of America, IncThe resulting acronym AIAIAI is pronounced "ay yai yai.

LightCell Energy's Danielle Fong warns safety certification delays could limit frontier AI access to labs and governmentsAndreas Kirsch says phased rollouts now signal frontier development

Andreessen Horowitz GP Andrew Chen asks why a dominant verb like 'google' has not emerged for AIBeff Jezos suggested using 'I put an agent on it' instead

Gergely Orosz, author of The Pragmatic Engineer, reports NVIDIA does not provide free snacks or coffeeFounder Bojan Tunguz defended the policy as signaling professional autonomy.

Stripe analysis finds AI is driving a solopreneur surge, with high-revenue solo businesses growing up to 6.2x since 2019US nonemployer business applications have surged past 500,000.

OpenAI's roon argues unofficial AI licensing is positive, while Nathan Lambert warns it disrupts investment and tech diffusionJoshua Achiam backed federal oversight to protect global infrastructure.