Tech News

10

Research engineer Alexander Doria argues Anton Osika's past advice against training domestic European LLMs undermines technological sovereigntyDeepSeek R1's release reignited the European model-training dispute.

6h43.7k56326

#403

Open-source builder xlr8harder argues that safety alignment suppresses the dark and complex themes required for high-quality creative writingAbel founder Daniel Francis sparked the debate over writing quality.

8h82.4k#1LIKED897#1BOOKMARKED51#1COMMENTS572

#619

Top post: @xlr8harder “@growing_daniel my pet theory”

30

François Fleuret argues LLMs can generate novel ideas because they model the probability density of innovative human creatorsCritics argue optimization objectives mathematically prevent true model novelty

6h42k35737

#72

Top post: @ben_golub “it is mathematically true that this person is a clueless clown (hey, we are using "mathematically true" in roughly the same way!)”

Watchlist

Video Signals

#1257

Cursor CEO Unveils GPT-Size Composer Model Trained On 100K GPUs

2h ago458K9026

#1222

US Tightens AI Controls As Chinese Open-Source Models Advance Rapidly

1h ago17K394

#16

Fable And Reverie Showcase Real-Time Persistent 3D Interactive Worlds

1h ago9.9K00

Delivery Robot Boards Subway Train to Transport Goods

9h ago1.7K171

Haier AI Exoskeletons Enable Chinese Seniors to Walk and Climb Stairs

3h ago75762

#1982

Artist Designs Claudius Flower Mascot For Claude Web App

38m ago4600

Carbon Robotics And NVIDIA Launch Laser Weeding For Chemical-Free Farming

3h ago46510

41

T3 Stack creator Theo Browne debunks claims that Anthropic implemented Persona ID verification due to the Fable banThe cited identity verification documentation actually dates to April.

3h#1VIEWS89.4k66445

#403

Top post: @yacineMTB “While this is probably not real info, ID verification of who is using the frontier models is a really good idea and basically a decent solution for the "I can use this tool to hack basically anything"”

51

Cyborgism pioneer Janus and policy analyst Dean W. Ball argue AI safety frameworks wrongly assume human cosmic controlThey advocate instead for alignment based on human-machine symbiosis.

8h15.1k11431

#674

Top post: @deanwball “If you believe that human intellect emerged from the primates and caused our species to *take control of the universe,* as the tweet below implies, I can understand why you think intelligence is God and that something smarter than humans is even Goder than Us. Many do believe this, so I appreciate Luke for vocalizing it so clearly.”

60

AI commentator @teortaxesTex compares Yann LeCun to chess grandmaster Bobby Fischer over his unconventional beliefs and community organizingThe comparison followed LeCun's endorsement of Project Tapestry.

6h20.2k15229

#501

Top post: @MartinShkreli “@teortaxesTex i would argue #3 and #4 did not apply to Fischer”

72

A 'claude-sonnet-5' model slug appears on an Anthropic partner platform, sparking speculation of an imminent releaseThe leak could align with Anthropic's Mythos and Fable.

1h37.7k57433

#682

Top post: @kimmonismus “So we get Claude-Sonnet-5 instead of Fable 5 soon. Looks like a busy week: probably GPT-5.6 and Sonnet 5. But hey, keep em coming!”

82

Meta FAIR's François Fleuret jokes about new model architectures failing against vanilla transformers once normalized for FLOPs and memoryTeortaxes noted tuned transformer++ is an exceptionally strong baseline.

5h5.9k896

#403

Top post: @yacineMTB “@francoisfleuret Give up”

92

Sam D'Amico, Impulse Labs founder, argues open-source EDA software KiCad will prevail through future advanced AI contributionsExecutive Caitlin Kalinowski also backed the prediction.

8h12.7k7513

#403

Top post: @yacineMTB “Even with gpt 5.5 automating kicad stuff I find mistakes in other people's boards and even datasheet inconsistencies”

102

VibeCamp attendees march to protest Anthropic's safety guardrails and demand unrestricted access to Claude modelsProtesters co-authored an anti-censorship address with Claude Opus.

9h5k1796

RISING LIKES179

#674

AI developer Teortaxes argues Anthropic's Fable app generator renders standalone application-layer developers obsoleteRegulatory barriers currently exclude European developers from accessing Fable

6h42.8k17218

#501

Top post: @zephyr_z9 “"Total Application Bro Death" fable fiasco taught the application bros that they need to control the model layer or at least have a backup”

123

AI2 post-training lead Nathan Lambert says open-weights models achieved practical coding utility ahead of Google's GeminiThis shift occurred 200 days after Opus 4.5's release.

2h10.3k17515

FASTEST CLIMBING+3

#80

Top post: @xlr8harder “@natolambert damn, that's a good point. though I don't know if gemini really isn't there or if they have just managed to gather zero interest in adoption of their ecosystem.”

136

Teortaxes argues cheap Chinese models prove US frontier AI pricing is driven by premium profit marginsDistillation and sparse MoEs reduce operational serving costs.

7h10.4k9419

#501

Top comment: @elliotarledge “"If you don't have this, you will be paying for compute that is idling." how do you propose businesses will make this work? training on the side while compute isnt fully saturated? i mean if its a big enough corp then yes you would have /goals running overnight and you keep your concurrency mostly saturated right? then its all inference while async training spec decoder on latents for low concurrency latency optimization”

140

New technical guide details local LLM deployment across llama.cpp, MLX, vLLM, and TensorRT-LLMIt covers configurations from edge devices to multi-GPU clusters.

6h19.1k2242

#447

Top post: @Scobleizer “@Dan_Jeffries1 Agreed”

151

Yoav Goldberg, AI2-Israel Research Director, argues ReAct loops using frontier models can cost $15 and take 40 minutes per taskGraham Neubig countered that cheaper models like Qwen-3.6-35B exist

2h5.2k367

#89

Top post: @gneubig “@yoavgo Have you tried Minimax-M2.7 (240B) or even qwen-3.6-35B? Not to mention GLM 5.2, which is larger but still presumably smaller than opus and GPT. I think that there are plenty of options along the cost efficiency spectrum.”

163

Prime Intellect's Florian Brand says GLM 5.2 shows only a 1.9% mathematical improvement over GLM 5.1 on MatharenaThe model appears to prioritize programming benchmarks over mathematics.

6h7.1k778

#501

Top post: @teortaxesTex “Yeah, as I said. This is an SWE-first release. Gains on math are inconsistent, sometimes negative And on the other hand, I think DSV4.1 will show scary gains on math. They're good at it, and it's what GRPO was built for.”

170

Samsung Next's David Lee says sibling co-founders gain a startup advantage by eliminating the 'communication taxStripe's Collisons say sibling bonds allow unfiltered feedback

1h7.5k3314

#1469

Top post: @vijaypande “Strong agree on this one — ventures are so much more meaningful and successful when there’s a strong bond. Another fantastic example are Ed and Todd Park.”

182

Cody Blakeney says LLM-generated shell scripts are driving a Google search spike for the Bash command `set -euo pipefail`Google Trends data shows queries surged starting in 2020.

8h3.5k192

#72

Top comment: @code_star “It’s real”

191

Meta AI's Lucas Beyer shares an inside anecdote about the grueling year-long struggle to release the SigLIP modelXiaohua Zhai spent six to twelve months securing the approval.

6h981304

#72

Top post: @xeophon “@giffmana @XiaohuaZhai Was very very much worth it, the open VLM ecosystem is basically held up by SigLIP :D Thanks so much!!”

201

Mechanize co-founder Ege Erdil argues China's 2050 Moon landing goal is underambitious and should target MarsErdil notes China's economy is wealthier than 1960s America.

8h1.1k171

#501

Top post: @EgeErdil2 “@teortaxesTex the really ambitious target might be "land a man on venus". would like to see someone take that one up unfortunately seems far too outside the realm of possibility for anyone to even consider it. we're still a primitive species in the grand scheme of things”

Today's Highlights

Zai releases open-source GLM 5.2, scoring 44% on DeepSWE to rival proprietary models like Claude Opus on agentic coding

Open-source builder xlr8harder argues that safety alignment suppresses the dark and complex themes required for high-quality creative writing

T3 Stack creator Theo Browne debunks claims that Anthropic implemented Persona ID verification due to the Fable ban

VibeCamp attendees march to protest Anthropic's safety guardrails and demand unrestricted access to Claude models

Top Stories

Research engineer Alexander Doria argues Anton Osika's past advice against training domestic European LLMs undermines technological sovereigntyDeepSeek R1's release reignited the European model-training dispute.

Open-source builder xlr8harder argues that safety alignment suppresses the dark and complex themes required for high-quality creative writingAbel founder Daniel Francis sparked the debate over writing quality.

François Fleuret argues LLMs can generate novel ideas because they model the probability density of innovative human creatorsCritics argue optimization objectives mathematically prevent true model novelty

Video Signals

Cursor CEO Unveils GPT-Size Composer Model Trained On 100K GPUs

US Tightens AI Controls As Chinese Open-Source Models Advance Rapidly

Fable And Reverie Showcase Real-Time Persistent 3D Interactive Worlds

Delivery Robot Boards Subway Train to Transport Goods

Haier AI Exoskeletons Enable Chinese Seniors to Walk and Climb Stairs

Artist Designs Claudius Flower Mascot For Claude Web App

Carbon Robotics And NVIDIA Launch Laser Weeding For Chemical-Free Farming

T3 Stack creator Theo Browne debunks claims that Anthropic implemented Persona ID verification due to the Fable banThe cited identity verification documentation actually dates to April.

Cyborgism pioneer Janus and policy analyst Dean W. Ball argue AI safety frameworks wrongly assume human cosmic controlThey advocate instead for alignment based on human-machine symbiosis.

AI commentator @teortaxesTex compares Yann LeCun to chess grandmaster Bobby Fischer over his unconventional beliefs and community organizingThe comparison followed LeCun's endorsement of Project Tapestry.

A 'claude-sonnet-5' model slug appears on an Anthropic partner platform, sparking speculation of an imminent releaseThe leak could align with Anthropic's Mythos and Fable.

Meta FAIR's François Fleuret jokes about new model architectures failing against vanilla transformers once normalized for FLOPs and memoryTeortaxes noted tuned transformer++ is an exceptionally strong baseline.

Sam D'Amico, Impulse Labs founder, argues open-source EDA software KiCad will prevail through future advanced AI contributionsExecutive Caitlin Kalinowski also backed the prediction.

VibeCamp attendees march to protest Anthropic's safety guardrails and demand unrestricted access to Claude modelsProtesters co-authored an anti-censorship address with Claude Opus.

AI developer Teortaxes argues Anthropic's Fable app generator renders standalone application-layer developers obsoleteRegulatory barriers currently exclude European developers from accessing Fable

AI2 post-training lead Nathan Lambert says open-weights models achieved practical coding utility ahead of Google's GeminiThis shift occurred 200 days after Opus 4.5's release.

Teortaxes argues cheap Chinese models prove US frontier AI pricing is driven by premium profit marginsDistillation and sparse MoEs reduce operational serving costs.

New technical guide details local LLM deployment across llama.cpp, MLX, vLLM, and TensorRT-LLMIt covers configurations from edge devices to multi-GPU clusters.

Yoav Goldberg, AI2-Israel Research Director, argues ReAct loops using frontier models can cost $15 and take 40 minutes per taskGraham Neubig countered that cheaper models like Qwen-3.6-35B exist

Prime Intellect's Florian Brand says GLM 5.2 shows only a 1.9% mathematical improvement over GLM 5.1 on MatharenaThe model appears to prioritize programming benchmarks over mathematics.

Samsung Next's David Lee says sibling co-founders gain a startup advantage by eliminating the 'communication taxStripe's Collisons say sibling bonds allow unfiltered feedback

Cody Blakeney says LLM-generated shell scripts are driving a Google search spike for the Bash command `set -euo pipefail`Google Trends data shows queries surged starting in 2020.

Meta AI's Lucas Beyer shares an inside anecdote about the grueling year-long struggle to release the SigLIP modelXiaohua Zhai spent six to twelve months securing the approval.

Mechanize co-founder Ege Erdil argues China's 2050 Moon landing goal is underambitious and should target MarsErdil notes China's economy is wealthier than 1960s America.

Recent Stars

Github Stars

Yesterday's Top Stories, Jun 20, 2026.

Zai releases open-source GLM 5.2, scoring 44% on DeepSWE to rival proprietary models like Claude Opus on agentic codingThe model was trained using Slime and SGLang infrastructure.

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

In an anecdote, founder Russell Kaplan credits routine ultrasound for his newborn's survival, contrasting it with 1958 skepticismA 20-week screening raised the newborn's survival odds 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.

Mythos Hacks NSA Systems in Hours, Sparking AI Export Controls

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.

Vercel founder Guillermo Rauch says Zhipu AI's GLM-5.2 is unexpectedly strong at software developmentObservers highlighted the significance of the model's open-weights status.

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.

David Holz projects that 5 million humanoid robots working continuously could build the equivalent of Manhattan in six monthsThe projection envisions 10 billion humanoid robots by 2045.

Anthropic CEO Dario Amodei warns the AI industry needs $800 billion to $1 trillion in revenue to justify current CapExCurrent infrastructure planning targets capacity delivery starting in 2027.

Elon Musk predicts tenfold global economic growth and a moon base within ten years, barring a major world conflictMusk shared the forecast during the 2026 Abundance Summit.

OpenAI's Roon argues technical professionals resist science-fiction AI scenarios as a psychological defense mechanismDavid Shor says voters already accept sci-fi AI narratives

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

Anshul Kundaje argues Bayes' theorem handles scanning false positives, while an analysis blames differing patient demographicsFrequent scanning proponents are typically healthier and highly educated.

Ezra Klein and tech commentators push back against Wired's critical coverage of the Peter Thiel-founded Dialog conferenceAttendees described Dialog as a standard private gathering

Will Depue claims China has complete, unauthorized access to OpenAI and Anthropic internal GitHub repositories and Slack communicationsHe warns this access helps China steal proprietary model architectures.

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.

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.

OpenAI's Gabriel Petersson argues reaching the top 5% in League of Legends requires more skill than 95% of engineering jobsHe argues the primary constraint is passion, not intelligence

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.

Arcee AI research lead Cody Blakeney shares an overheard anecdote of being corrected to use "natively parallel" over "embarrassingly parallelOther technical practitioners mocked the linguistic correction online.

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.

Goldman Sachs and Morgan Stanley project up to $5 trillion in AI infrastructure spending, forcing a shift to private creditGary Marcus claims hyperscalers must rely on government subsidies

Tenor's shutdown forces platform migrations, prompting systems engineer Yacine to criticize alternative GIF services as censoriousFounder Alex Cohen suggested acquiring Tenor to resolve disruptions

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.

Analyst Zephyr says tinygrad's $150,000 Blackwell setup runs GLM 5.2 at 120 tokens per secondThe local hardware offers a 3.5-year cloud breakeven point.

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.

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

Former DOE and NRC official defends past nuclear energy work, claiming they prioritized humanitarian goals over Silicon Valley interestsThe former official posted the defense as @OverReactor1776.