Tech News

10

Nabeel S. Qureshi says a Harper’s Bazaar prize-winning short story was flagged as 100% AI-generated by PangramHe urged contests to verify submissions or permit AI.

10h567.9k2.2k723

#184

Top post: @__nmca__ “the fact that slop has won a couple of literary prizes implies that slop is in large part an exposure effect; more broadly frequent AI users probably underestimate how good slop is by 2019’s standards. (story is ass imo)”

Digg Deeper question: “What's pangram?”

answerview answer

20

30

Watchlist

Video Signals

#792

TikTok Users Deploy AI to Create Fake Early 2000s Actresses

4h ago188K667458

#1861

Brian Roemmele Debunks Myth Superintelligent AI Will Hate Humans

2h ago16K8926

#192

Claude Generates Psychological Profile Of Anthropic CEO Dario Amodei

1h ago39K488140

#1360

Trump Says Anthropic No Longer Poses National Security Threat

10h ago35K33352

#627

Levelsio Adds AI Chat Assistant And Anti-ESG Filter To Hotelist.com

3h ago70K342205

Amazon Drops Sam Altman Film 'Artificial' Starring Andrew Garfield

5h ago148K16624

#1407

Dean Ball Urges Broad AI Model Diffusion to Avert Nationalization Risks

9h ago11K7427

#447

Scoble Explores AI Governments Inspired By Edge Esmeralda Builders Event

10h ago6.2K3010

#1222

Podcast Examines Hyperscaler Weakness And Rise Of Open Source AI Models

8h ago20K10133

#33

S-Agent Uses Spatial Tool-Use To Boost Reasoning On Vision Benchmarks

7h ago6.9K268

40

50

Abridge co-founder Zachary Lipton jokes that Barret Zoph is joining Thinky Machines as CFO, satirizing rapid AI talent movesDivyansh Kaushik suggested restarting the Approximately Correct publication.

7h375.2k1.2k143

#31

Yunta Tsai's analysis finds machine learning projects spend 50% of effort on evaluation and only 2% on trainingData cleaning takes 40%, a breakdown Elon Musk recently endorsed.

6h#1VIEWS9.4M#1LIKED24.2k#1BOOKMARKED2.3k#1COMMENTS1.8k

#52

Top comment: @mahfuzxallum “@elonmusk Data quality is still the highest leverage optimization in ML”

Digg Deeper question: “Sheing Ng @sashimikun_void how many percentage of that 2% training is run by AI? educated guess”

answerview answer

74

Stability AI co-founder Emad Mostaque teases that an open-source 'fable-level' AI model will run on base MacBook hardwareHe provided no technical specifications, architecture details, or timeline

7h124.7k2k197

#171

Top post: @EMostaque “My guess is 18 months fwiw. My forecasts have been reasonably accurate on capabilities previously.”

84

Meta FAIR's François Fleuret argues speculative decoding is a nearly free lunch, prompting debate over token volume versus deeper reasoningResearch engineer kache argued systems should focus on deeper thought.

10h31.5k342118

#149

Top post: @yacineMTB “@francoisfleuret why grug make token. grug think. less token. more thought”

94

Open-source GLM 5.2 beats Claude Opus 4.8 on the Official Instruct Models benchmark, prompting one Fortune 500 firm to drop AnthropicObservers question GLM 5.2's performance in real-world agent loops.

8h60.6k582236

#501

Top post: @AndrewCurran_ “I've been posting extremely positive reviews of GLM 5.2 here, but some people are also having the complete opposite reaction. This was also exactly the same way GPT-5 and Opus 4 were initially received. The review chart looked like this \_/ with huge tails on either end.”

101

THUDM open-sources Slime, an LLM post-training framework for reinforcement learning scaling and online preference optimizationIt completed GLM-5.2 post-training in two days.

14h6.9k12886

#90

Top post: @suzatweet “this is actually really cool, fully open source. RL framework behind GLM-5.2, GLM-5.1, GLM-5, GLM-4.7, GLM-4.6, and GLM-4.5, validating the full post-training loop.”

114

Debate centers on Midjourney scanner's lower imaging quality versus potential for billion-person diagnostic datasetsEngineers note speed enables 10-1000x more scans for AI training.

10h42.8k52738

#1315

Top post: @NickADobos “The data from midjourney scanning 1 billion people will be more important than the scanner itself or the initial diagnostic potential The scanner could be less effective than other modern medical imaging, doesn't really matter, think about this like a data scientist & business not a doctor. Putting it in a spa you can walk into in 60 seconds fundamentally changes how many body scans you can harvest by 10-1000x. The diagnostic and research potential of that data is astounding. What does a person's medical history & treatment plan look like if they have full body scans every week? They are building a spa that sells you your body's data.”

Digg Deeper question: “It has been said that this type of technology cannot penetrate bone and therefore could not be used to see inside the brain or chest area, is this true?”

answerview answer

126

Geoffrey Litt argues AI coding agents struggle with UI because specification and implementation tasks constantly blurStrongDM's Attractor spec proposes non-interactive loops to ease delegation.

5h47.3k426402

#1402

Top post: @jamescham “@andy_matuschak Justin McCarthy’s attractor spec gives some detailed perspective on how a longer loop could be run. https://github.com/strongdm/attractor”

136

Daniel Francis, Abel founder, criticizes San Francisco's rise of manufactured tech events used primarily for self-promotionKylie Robison agreed, citing a surge in forced AI events.

15h233.1k3.2k323

#619

Top comment: @SimonTheNoob “@growing_daniel Honestly, I think this is just chinese honeypot, with hot chinese chicks trying to get information from founders or employees of certain companies. Just keep your mouth shut and enjoy the company.”

146

Amazon MGM Studios drops Luca Guadagnino's OpenAI coup movie 'Artificial' starring Andrew Garfield as Sam AltmanTest audiences disliked the Sam Altman and Elon Musk characters.

9h180.3k36159

#178

Top post: @TrungTPhan “a notable detail of Andrew Garfield acting as Sam Altman is that they can just re-use this scene from “The Social Network” for when Sam finds out about Zuck poaching OpenAI researchers with $100 million offers”

156

Miles Brundage and Susan Zhang argue against nationality-based restrictions on AI talent, urging a focus on basic securityThe debate focused on national security and corporate espionage.

15h12.2k1049

#57

Top post: @suchenzang “oh missed this one. unless you have hard proof that your former colleagues are compromised, have leaked secrets, and have openai legal team pursuing the issue with the FBI, i suggest you stop pretending that what you're saying isn't sinophobia by definition. keep it a technical issue. or keep race out of it. slapping on a few other countries after starting out with "china is stealing" does not help your case.”

1610

Joseph Barrow and Denis Peskoff of UC Berkeley release a dataset of 2.2 million US city and county lawsThe corpus aggregates previously fragmented local laws for public search

7h8.3k189

#33

Top comment: @barrowjoseph “Once we had processed the laws, the question was: how to analyze them. One idea I really like is @TheStalwart's http://Havelock.ai, where he distills pairwise comparisons from an LLM + TrueSkill into a BERT model to score orality.”

176

OpenAI's roon predicts a 10,000-fold increase in personal data logging, with AI filtering continuous lifetime trackingInvestor Josh Wolfe endorsed the premise, calling it lifecording.

4h58.7k619136

#44

Top post: @wolfejosh “100%. LIFECORDING”

184

AI Alliance launches Project Tapestry to co-train open frontier models and counter Chinese AI dominanceParticipants retain sovereign derivatives while sharing a base model.

12h25.6k13663

#3

Top post: @Dorialexander “Happy to be onboard.”

196

Developer @tensorqt sets a new modded-NanoGPT speedrun record of 2,925 steps to reach 3.2781 validation lossParameter-space extrapolation shaved five steps off the previous record.

4h13.2k13730

FASTEST CLIMBING+6

#703

Top post: @kellerjordan0 “New Modded-NanoGPT optimization SOTA: @tensorqt has achieved a 2925-step run (-5 steps vs. prev SOTA) by adding a late parameter-space extrapolation step to the previous record.”

206

Z.ai launches GLM-5.2, a coding agent trained on SGLang and Slime that reportedly matches Claude Opus capabilitiesThe system builds websites and manages long-horizon tasks.

14h24.5k24251

#27

Top post: @IntuitMachine “If GLM-5.2 has Opus 4.8-like capabilities, then what prevents http://Z.ai from creating a Mythos/Fable level AI? Is the cat already out of the bag wrt higher capable models?”

Today's Highlights

Nobel laureate and AlphaFold lead John Jumper departs Google DeepMind to join rival AI lab Anthropic

Yunta Tsai's analysis finds machine learning projects spend 50% of effort on evaluation and only 2% on training

Developer @tensorqt sets a new modded-NanoGPT speedrun record of 2,925 steps to reach 3.2781 validation loss

Lux Capital co-founder Josh Wolfe questions Elizabeth Warren's proposal to tax Elon Musk to fund free community college

Top Stories

Nabeel S. Qureshi says a Harper’s Bazaar prize-winning short story was flagged as 100% AI-generated by PangramHe urged contests to verify submissions or permit AI.

Menlo Ventures partner Deedy Das argues AI coding assistants are splitting developers into automators and code-quality managersDiligent engineers face burnout from correcting automated code output

Parody website InTheWeights.com launches to rank tech figures on their hypothetical presence in fictional future AI model weightsAlexis Ohanian and Shawn Wang shared their mock profiles.

Video Signals

Russell Kaplan, Cognition AI co-founder, highlights how early obstetric ultrasound faced intense skepticism before becoming a life-saving standardDetecting vasa praevia before birth raises fetal survival to 99%.

Abridge co-founder Zachary Lipton jokes that Barret Zoph is joining Thinky Machines as CFO, satirizing rapid AI talent movesDivyansh Kaushik suggested restarting the Approximately Correct publication.

Yunta Tsai's analysis finds machine learning projects spend 50% of effort on evaluation and only 2% on trainingData cleaning takes 40%, a breakdown Elon Musk recently endorsed.

Stability AI co-founder Emad Mostaque teases that an open-source 'fable-level' AI model will run on base MacBook hardwareHe provided no technical specifications, architecture details, or timeline

Meta FAIR's François Fleuret argues speculative decoding is a nearly free lunch, prompting debate over token volume versus deeper reasoningResearch engineer kache argued systems should focus on deeper thought.

Open-source GLM 5.2 beats Claude Opus 4.8 on the Official Instruct Models benchmark, prompting one Fortune 500 firm to drop AnthropicObservers question GLM 5.2's performance in real-world agent loops.

THUDM open-sources Slime, an LLM post-training framework for reinforcement learning scaling and online preference optimizationIt completed GLM-5.2 post-training in two days.

Debate centers on Midjourney scanner's lower imaging quality versus potential for billion-person diagnostic datasetsEngineers note speed enables 10-1000x more scans for AI training.

Geoffrey Litt argues AI coding agents struggle with UI because specification and implementation tasks constantly blurStrongDM's Attractor spec proposes non-interactive loops to ease delegation.

Daniel Francis, Abel founder, criticizes San Francisco's rise of manufactured tech events used primarily for self-promotionKylie Robison agreed, citing a surge in forced AI events.

Amazon MGM Studios drops Luca Guadagnino's OpenAI coup movie 'Artificial' starring Andrew Garfield as Sam AltmanTest audiences disliked the Sam Altman and Elon Musk characters.

Miles Brundage and Susan Zhang argue against nationality-based restrictions on AI talent, urging a focus on basic securityThe debate focused on national security and corporate espionage.

Joseph Barrow and Denis Peskoff of UC Berkeley release a dataset of 2.2 million US city and county lawsThe corpus aggregates previously fragmented local laws for public search

OpenAI's roon predicts a 10,000-fold increase in personal data logging, with AI filtering continuous lifetime trackingInvestor Josh Wolfe endorsed the premise, calling it lifecording.

AI Alliance launches Project Tapestry to co-train open frontier models and counter Chinese AI dominanceParticipants retain sovereign derivatives while sharing a base model.

Developer @tensorqt sets a new modded-NanoGPT speedrun record of 2,925 steps to reach 3.2781 validation lossParameter-space extrapolation shaved five steps off the previous record.

Z.ai launches GLM-5.2, a coding agent trained on SGLang and Slime that reportedly matches Claude Opus capabilitiesThe system builds websites and manages long-horizon tasks.

Recent Stars

Github Stars

Yesterday's Top Stories, Jun 19, 2026.

Nobel laureate and AlphaFold lead John Jumper departs Google DeepMind to join rival AI lab AnthropicDeepMind CEO Demis Hassabis praised Jumper's historic contributions.

Elon Musk argues interstellar travel will shift economic metrics from currency to mass and energyCasey Handmer outlined a roadmap to scale antimatter production.

Will Depue, former OpenAI Sora researcher, sparks debate over claims that Chinese espionage compromised OpenAI and Anthropic systemsNathan Lambert argued talent movement spreads technical knowledge anyway.

US Commerce Department orders Anthropic to get approval before letting foreign nationals access its Fable 5 and Mythos 5 modelsTrump no longer considers Anthropic a national security threat.

US Commerce Secretary Howard Lutnick alleges an ASML EUV lithography machine reached China, which ASML deniesChinese research teams are actively developing domestic EUV prototypes

Thomas Dimson launches In The Weights, a retro site visualizing how strongly a name is represented in AI training dataThe search tool generates video game-style character profile cards

Roon argues human-computer hybrids cannot compete with pure machine intelligence due to slow biological brainsCommentators counter with metrics like energy efficiency and consciousness.

Codex launches automatic session handoff to transfer active developer threads between local laptops and remote serversUsers can close local laptops without losing active progress

OpenHands co-founder Graham Neubig says GLM-5.2 is the first open model capable of replacing closed-source models in his coding workflowsThe 744B model features a one-million-token context window

Leaked Google DeepMind report reveals internal discontent as its AI models slip to fifth place on industry benchmarksEmployees worry leadership is conceding the race to AGI

Boris Cherny claims to crack the undeciphered 3,500-year-old Linear A script using Anthropic's Claude CodeThe translation attempts will undergo academic peer review.

Dwarkesh Patel argues the massive sample efficiency gap between LLMs and humans represents a critical AI data bottleneckLLMs require up to a million times more training data.

Proponents claim Midjourney's low-cost whole-body scanner overcomes diagnostic limits, but critics argue hardware upgrades cannot solve clinical failuresProponents claim the technology cuts imaging costs by 99.9%.

Midjourney launches a medical division, unveiling a conceptual 60-second full-body ultrasonic scanner to replace MRIsMidjourney plans to deploy 50,000 scanner units globally.

Analysis argues China's GLM challenges data-scarcity assumptions by matching Claude 3 Opus performance through model distillationSanctions forced labs to use consumer GPUs for RL.

Figure founder Brett Adcock projects the company's humanoid robot fleet will surpass its human headcount by Q2 2026The projection targets 700 robot units against 650 employees.

Policy researcher Helen Toner says Anthropic rapidly builds advanced AI to map existential risks before rivals trigger themShe compared scaling AI to mapping a monster-filled forest.

THUDM open-sources Slime, the reinforcement learning scaling framework used to train Zhipu AI's GLM-5.2 modelIt merged over ten expert models in two days

Jean Kaddour releases Sokoban Speedrun, an RL benchmark that fine-tunes Qwen3-4B-Instruct in 87 minutes using GRPOThe project modifies Andrej Karpathy's open-source nanochat pipeline

EU issues tenders for seven data centers with up to 40,000 GPUs, scaling back sovereign infrastructure plansObservers say the downsized footprint favors smaller AI models.

Ukraine launches TrophyLab, a secure platform for sharing vulnerability analyses and data on captured Russian military hardwareAuthorized partners can also request physical weapon samples

Mitchell Hashimoto, Ghostty creator, embeds prompt injections in codebases to detect and ban contributors submitting unreviewed AI-generated codeThe traps are placed in files like AGENTS.md.

Gabriel Petersson, who worked on OpenAI's Sora, confirms he left the company a month before the official announcementAn outside user leaked his departure weeks before the announcement.

Stripe's Patrick Collison says the Tempo blockchain hit a $3 billion annualized payment volume run rate 93 days after launchThe network hosts 1,000 payment services for AI agents.

Study of 26,811 students finds generative AI cuts homework time but reduces exam scores by 20%AI hurts learning when it replaces active mental effort.

Pirate Wires report says Medium burned up to $2.6 million a month while losing writers to SubstackMedium raised $157 million before tightening content moderation

Noah Smith parodies formulaic tech career announcements, mocking rapid shifts between hyped sectorsAntonio García Martínez called these updates an insufferable trope.

White House and Anthropic collaborate on a standardized technical framework to benchmark AI jailbreak severityThe framework could restore access to Mythos and Fable.

Audit of Google DeepMind's DiffusionGemma finds intermediate steps remain human-interpretable, enabling chain-of-thought safety monitoringThis restores transparency benefits previously restricted to autoregressive LLMs.

Systems engineer Yacine argues tech companies should stop hiring from NeurIPS and CVPR and recruit from SIGGRAPH insteadHeinrich Kuttler noted DeepMind's original team relied on gaming.