Tech News

Today's Highlights

Midjourney launches Midjourney Medical and unveils a conceptual ultrasound scanner designed to perform full-body scans in 60 seconds

#52

+45 more

It’s a new day at Digg. Fresh stories are clustering now. 55% complete.

Top Stories

Posts:—Clusters:—Next Crawl:—

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

4h10.8k15764

#501

Top comment: @teortaxesTex “@didier_lopes > The entire OPD post-training of GLM-5.2 took on this slime platform took ~2 days. what”

Jürgen Schmidhuber argues his Munich team's 1991 papers introduced the foundational concepts for modern transformers, network distillation, and pre-trainingSakana AI's hardmaru says the work influenced his research.

3h20k17741

#17

Top post: @hardmaru “In 1991, the foundations for Transformers, Pre-training, Distillation, and World Models were already being built. These helped shape my own thinking, from my time at Google Brain to our Recursive Self-Improvement (RSI) work at @SakanaAILabs today. 🧠🗼 https://people.idsia.ch/~juergen/ai-boom-roots-munich-1991.html 👇”

Teortaxes argues Zhipu AI and Tsinghua University will build a competitive Chinese AI ecosystem within a yearOpen-source contributor Stella Biderman agreed with the assessment.

3h12.4k10910

#289

Top post: @teortaxesTex “Come to think of it, Zhipu vs DeepSeek is pretty much Tsinghua academics vs Peking University competitive programmers. Whom would you ask to do the best research? Whom to design the most performant algorithms? Of course they're all great researchers and engineers, but…”

Watchlist

Video Signals

#1360

WoW Private Server Runs 1,800 AI Bots With DeepSeek Chat

34m ago3.8K6112

#809

Crysta AI 3D Prints Gaussian Splat Bee Model in Resin Cube

2h ago69K00

Google Gemini 3.1 Flash TTS Adds Actor-Level Voice Direction Controls

2h ago29552

China's 24/7 AI-Run Textile Factory Sets New Standard for Autonomous Manufacturing

1h ago24410

Lentil Robotics Cuts Robot Vision Setup Time from Weeks to 6 Minutes

4h ago40300

Boston Dynamics Integrates Gemini AI into Spot Robot for Autonomous Tasks

10m ago7800

Noam Shazeer's departure from Google raises doubts about Gemini Pro's future as OpenAI recruits pre-training talentOpenAI is hiring London-based pre-training researchers for model scaling

2h8.4k10025

#1498

Top post: @bindureddy “Noam Shazeer leaving Google doesn’t bode well Hopefully Gemini will continue to be a credible model family Flash is a good model but Pro is still up in the air 🤷‍♀️”

Elie Bakouch and Meta's Lucas Beyer argue a hypothetical AI scaling challenge would reward pre-existing infrastructure over pure capabilityTeams must use existing codebases instead of starting fresh

3h99670

#72

Top post: @matthen2 “@eliebakouch google people would just complain about not having TPUs 😅”

High-achieving South Korean engineering students increasingly choose elite Chinese universities over domestic optionsThe shift reverses historical patterns of China recruiting Korean talent

2h3.4k376

#501

Top post: @davidlee “If true, this would be a great trend for Korea. Exposing the best young talent to a different mindset towards engineering, building, experimentation washes away the value that failure is shame”

Resurfaced demo shows Midjourney founder David Holz personally coded a realtime browser-based 3D visualizationThe system runs 3D visualizations without external plug-ins.

4h2.6k343

#447

Top post: @Scobleizer “@olivercameron Totally!”

GLM 5.2 (max) scores 70.1% on WeirdML benchmark, requiring 22,000 output tokens per runA lower configuration yields 67.3% accuracy using 12,000 tokens.

1h4.2k321

#501

Top post: @teortaxesTex “off by 0.1%. I'm getting rusty Yes, 7 months of a gap. Of course GLM 5.2 is a better model than Gemini 3 Pro in most real use cases. WeirdML is a benchmark which profoundly separates Western and Eastern training culture… or used to.”

Google DeepMind's Susan Zhang highlights DeepSeek-R1's $600 billion Nvidia hit, while Will Brown points to long-term growthThe release triggered a 17 percent single-day stock decline.

4h1.1k202

#84

Top post: @willccbb “@suchenzang one rounding error among many”

101

Google DeepMind's Andreas Kirsch argues AI researchers can no longer ignore the downstream consequences of advancing model capabilitiesKirsch cites a lack of independent oversight and opaque deals.

1h63591

#271

Top post: @genekogan “@Miles_Brundage It doesn’t seem like many people work on capability and safety at the same time. I kind of think most people understand they are not the difference maker on whether the consequences turn out good or bad. If they genuinely think it’s the latter, they stop working on capability.”

111

Moondream AI co-founder Vikhyat K. says GLM-5.2's two-day OPD post-training works by collapsing multiple expert variantsThis method concentrates and densifies the model's training signal

3h622150

#501

Top post: @teortaxesTex “@vikhyatk @didier_lopes yeah the word "entire" was throwing me off but for OPD I guess makes sense, if they encounter no problems”

Recent Stars

Github Stars

(7 days)

Today's Highlights

Midjourney launches Midjourney Medical and unveils a conceptual ultrasound scanner designed to perform full-body scans in 60 seconds

Top Stories

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

Jürgen Schmidhuber argues his Munich team's 1991 papers introduced the foundational concepts for modern transformers, network distillation, and pre-trainingSakana AI's hardmaru says the work influenced his research.

Teortaxes argues Zhipu AI and Tsinghua University will build a competitive Chinese AI ecosystem within a yearOpen-source contributor Stella Biderman agreed with the assessment.

Video Signals

Noam Shazeer's departure from Google raises doubts about Gemini Pro's future as OpenAI recruits pre-training talentOpenAI is hiring London-based pre-training researchers for model scaling

Elie Bakouch and Meta's Lucas Beyer argue a hypothetical AI scaling challenge would reward pre-existing infrastructure over pure capabilityTeams must use existing codebases instead of starting fresh

High-achieving South Korean engineering students increasingly choose elite Chinese universities over domestic optionsThe shift reverses historical patterns of China recruiting Korean talent

Resurfaced demo shows Midjourney founder David Holz personally coded a realtime browser-based 3D visualizationThe system runs 3D visualizations without external plug-ins.

GLM 5.2 (max) scores 70.1% on WeirdML benchmark, requiring 22,000 output tokens per runA lower configuration yields 67.3% accuracy using 12,000 tokens.

Google DeepMind's Susan Zhang highlights DeepSeek-R1's $600 billion Nvidia hit, while Will Brown points to long-term growthThe release triggered a 17 percent single-day stock decline.

Google DeepMind's Andreas Kirsch argues AI researchers can no longer ignore the downstream consequences of advancing model capabilitiesKirsch cites a lack of independent oversight and opaque deals.

Moondream AI co-founder Vikhyat K. says GLM-5.2's two-day OPD post-training works by collapsing multiple expert variantsThis method concentrates and densifies the model's training signal

Recent Stars

Github Stars

Today's Highlights

Midjourney launches Midjourney Medical and unveils a conceptual ultrasound scanner designed to perform full-body scans in 60 seconds

Top Stories

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

Jürgen Schmidhuber argues his Munich team's 1991 papers introduced the foundational concepts for modern transformers, network distillation, and pre-trainingSakana AI's hardmaru says the work influenced his research.

Teortaxes argues Zhipu AI and Tsinghua University will build a competitive Chinese AI ecosystem within a yearOpen-source contributor Stella Biderman agreed with the assessment.

Video Signals

Noam Shazeer's departure from Google raises doubts about Gemini Pro's future as OpenAI recruits pre-training talentOpenAI is hiring London-based pre-training researchers for model scaling

Elie Bakouch and Meta's Lucas Beyer argue a hypothetical AI scaling challenge would reward pre-existing infrastructure over pure capabilityTeams must use existing codebases instead of starting fresh

High-achieving South Korean engineering students increasingly choose elite Chinese universities over domestic optionsThe shift reverses historical patterns of China recruiting Korean talent

Resurfaced demo shows Midjourney founder David Holz personally coded a realtime browser-based 3D visualizationThe system runs 3D visualizations without external plug-ins.

GLM 5.2 (max) scores 70.1% on WeirdML benchmark, requiring 22,000 output tokens per runA lower configuration yields 67.3% accuracy using 12,000 tokens.

Google DeepMind's Susan Zhang highlights DeepSeek-R1's $600 billion Nvidia hit, while Will Brown points to long-term growthThe release triggered a 17 percent single-day stock decline.

Google DeepMind's Andreas Kirsch argues AI researchers can no longer ignore the downstream consequences of advancing model capabilitiesKirsch cites a lack of independent oversight and opaque deals.

Moondream AI co-founder Vikhyat K. says GLM-5.2's two-day OPD post-training works by collapsing multiple expert variantsThis method concentrates and densifies the model's training signal

Recent Stars

Github Stars

Yesterday's Top Stories, Jun 18, 2026.

Midjourney launches Midjourney Medical and unveils a conceptual ultrasound scanner designed to perform full-body scans in 60 secondsMidjourney plans a San Francisco spa launch by 2027

OpenAI hires former White House AI adviser Dean W. Ball to lead its new Strategic Futures policy teamHe previously authored the White House AI Action Plan.

Instagram algorithm author Thomas Dimson launches In The Weights to check if your name is encoded in AI modelsThe tool measures name recognition across 14 different AI models.

Transformer co-author Noam Shazeer leaves Google to join OpenAI as lead of architecture researchHe also pioneered foundational research into mixture-of-experts systems.

Z.ai releases GLM-5.2, an open-weight model with a 1-million-token context window that outperforms GPT-5.5 on FrontierSWEThe model is distributed under an MIT license.

OpenAI launches Record and Replay for Codex, allowing users to turn manual tasks into reusable, editable skillsIt automates repetitive administrative tasks like expense reports.

Senator Bernie Sanders proposes giving the public a 50% equity stake in AI companies exceeding $200 million in revenueTech investors warned the mandate would deter new startups

OpenAI and Boston Children's Hospital use o3 Deep Research to diagnose 18 unsolved pediatric genetic casesThe findings were published in the journal NEJM AI.

Claude Code releases Artifacts beta to convert active coding sessions into interactive, shareable HTML pagesThe beta is exclusive to Team and Enterprise plans.

GLM creator Jie Tang disputes Elon Musk's Q1 2027 timeline for Chinese AI to reach Western frontier parityBenchmarks place GLM 5.2 seven months behind Claude.

US Vice President JD Vance warns that AI functions as a communist technology by enabling deep corporate and government surveillanceHe warned the technology could enable social credit systems

Coinbase CEO Brian Armstrong and Elon Musk argue orbital datacenters will surpass terrestrial compute to bypass Earth's regulatory bottlenecksArmstrong proposed capping regulatory growth in Mars legal frameworks

OpenAI's @tszzl questions the competitive risk of disclosing reinforcement learning and chain-of-thought scaling principles during the o1 launchThe debate highlights the competitive cost of publishing AI research.

US Commerce Department warns ASML that a restricted EUV lithography machine may have reached ChinaASML is the sole global producer of EUV systems.

Forward Future founder Matthew Berman launches Loop Library, a directory of repeatable, verified AI agent workflowsWorkflows include automated ticket-to-PR and page-load optimization loops

xAI integrates Grok models into Databricks Agent Bricks to let enterprises build AI agents on proprietary dataThe integration maintains secure enterprise control over private data.

Trump administration issues 90-minute ultimatum to Anthropic, forcing global suspension of Fable 5 and Mythos 5 modelsThe order bars non-US employees from training the models.

Anthropic Focuses on Useful AI Intelligence to Drive Revenue

The White House and Anthropic are developing a technical framework to assess AI jailbreak severity and establish security rulesThe negotiations shift US policy away from export controls.

Paul Graham says the next major computing form factor is unknown but will seem obvious in retrospectMinh Nhat Nguyen mocked the prediction with a meme.

A resurfaced post shows Midjourney co-founder David Holz proposed launching a biomedical imaging division to reduce scan false positivesThe March 2024 proposal targeted annual cancer scans.

Anthropic's Chris Ciauri says blocked frontier models Mythos and Fable 5 will regain availability in South Korea within daysAccess restoration follows positive G7 negotiations with President Trump

Accenture shares plunge up to 20% after lowered revenue forecast blames AI disruption of consulting workflowsThe selloff dragged down rivals Infosys, Tata, and Cognizant

Open-source builder xjdr launches Noumena Code to manage AI coding agents and resolve Git workflow bottlenecksIt includes a cloud workspace and terminal UI.

a16z's independent media division releases its first annual report detailing its direct-to-audience content strategy for foundersThe division operates separately from the core venture firm.

Richard Ngo argues the AI safety community's narrow focus on AGI and recursive self-improvement constrains thinking at labsLeo Gao countered that AGI concern is still essential.

Valar Atomics takes its Ward 250 advanced nuclear reactor critical for the first timeThe startup plans to deliver power by July 4.

Founders Fund releases final "MAFIA" episode featuring Sam Altman, Elon Musk, and Peter Thiel playing the social deduction gameThe 24-minute production features group discussions and card reveals

Poolside AI releases Laguna M.1, a 225B-A23B open-weights MoE model with a 256K context windowThe model is licensed Apache 2.0 for codebase exploration.

OpenAI releases beneficial reinforcement learning research evaluating GPT-5.5 and Claude Opus 4.7 on persistent alignmentThe framework tests how models maintain safe behavior under pressure.