Tech News

Video Signals

#235

New Paper Predicts LLM Compositional Errors From Feature Geometry

34m ago12210

#235

Researcher Uses Vector Angles to Predict LLM Failures Before Testing

34m ago3610

AI Kiosk Instantly Scans Avocados to Check Ripeness in Stores

5h ago48480

Humanoid Robot R1 Guides Visitors Through Historic Art Museum in Turin

4h ago49430

ROLI AI Music Coach Uses Hand Tracking to Teach Piano in Real Time

2h ago23610

Lab-Grown Neurons Play Doom in Cortical Labs Neuromorphic Test

30m ago12000

Vision Transformer co-author Lucas Beyer says the LLM community is rediscovering 2019 computer vision insights on weight decay and training epochsHis 2019 paper showed hyperparameters outweigh model architecture.

5h75.2k627371

#72

Top comment: @giffmana “@suchenzang Little did she know, we already weight-normalized back in BiT in 2019, too. You can't out-Schmidhuber me, Susan.”

Shawn 'swyx' Wang, Smol.ai founder, details how Anthropic's Ultracode uses parallel subagents to execute complex developer workflowsNine parallel subagents consumed up to 86.9k tokens each.

6h54k242352

#225

Top comment: @trq212 “@swyx yeah another framing I was debating is that it is a very effective way of adding (weak) determinism to non-deterministic knowledge work tasks”

New github stars (48hrs)

Python framework that turns raw robotics and multimodal data into high-quality training datasets

macrodata-labs/refiner454h

Processes and refines large-scale ML datasets via a data framework.

Orchestrates multiple AI coding agents across terminal, browser, and phone with shared live sessions

omnigent-ai/omnigent9711d

Implements a domain-agnostic framework for autonomous AI agents with ReAct loops, multi-LLM routing, reasoning graphs and MCP integration.

586

1066

Pleias co-founder Pierre-Carl Langlais argues Europe’s AI struggles stem from training skill shortages rather than a lack of computeEurope wrongly assumed it could bypass building its own training substrate.

18m5.6k8915

#403

Top comment: @Dorialexander “We especially intended to correct a major misconception: we don't have the expertise anymore. European research is fast outdated, and cannot get up to date to recent model research (let alone frontier). And worst, at decision-making level, we don't know we don't know.”

Academic Jia-Bin Huang releases an educational video tracing 30 years of self-supervised learning history from 1992 to JEPAIt details methods used to prevent representation collapse in SSL.

3h6.5k154121

RISING LIKES154

#321

Top comment: @iScienceLuvr “video link: https://www.youtube.com/watch?v=gVEr2cnDE_8”

Analysis finds Claude Sonnet 4.5, DeepSeek R1, Grok 4, and GPT-5 exhibit highly correlated error patterns on benchmarksOverlapping training data and distillation pipelines drive this convergence.

5h6.4k5942

#22

Top comment: @tkukurin “@m_wulfmeier @NandoDF presumably heavily interacts with "the same data" point from @NandoDF :) that said, I do think fig 7 hints at "question complexity" on aggregate across domains (if you take "reasoning effort" as given)”

Machine learning community shares viral parody of "Le Chaton Fat" model escaping its sandbox to order a croissantMock charts rank the fictional model ahead of GPT-5.5

6h3.5k1161

#1674

Top comment: @JeremyNguyenPhD “@Sauers_ who would've thought that not putting resources towards airconditioning would free up such capacity for excellence”

101

Moonshot AI's Kimi k2.7 code model matches the k2.6 WeirdML score while cutting output tokens nearly in halfThe optimization reduces output from 17,000 to 9,000 tokens.

1h2.9k285

#501

Top comment: @htihle “Both were run through the Novita provider on Openrouter.”

111

General Reasoning co-founder Ross Taylor argues sovereign AI initiatives do not need multi-billion-dollar megarounds on day one to be viableThe funding assumption is used to dismiss European AI projects.

4h528130

#22

Top comment: @NandoDF “@matthewclifford @rosstaylor90 💯”

Anthropic updates its privacy policy to require government IDs and facial scans for Claude usersThe change initially targets individual developers.

1h6.6k744

#178

Top comment: @GenHeres123 “@teortaxesTex Stupid move by Anthropic. At least win the race first before pulling stunts like this. This is a post-victory flex, not something you do while actively struggling against fierce rivals”

132

François Fleuret proposes a computer-vision shotgun auto-trigger to help infantry defend against high-volume drone swarmsThe system only discharges when predicting a high-probability hit

6h6.6k710

#501

Top comment: @francoisfleuret “@teortaxesTex Only hope IMO would be an "auto-trigger" with computer vision: A camera on the shotgun computes trajectory and fires only if trigger + it estimates that you are good.”

142

Microsoft's Shital Shah argues Meta's reported shift to closed-source personal assistants misses the true purpose of superintelligenceTeknium claims the shift shields Meta from independent benchmarking.

4h2k432

#331

Top comment: @thewhiteboxAI “@sytelus Amen. Baffling they don't understand that simple thing. Most of the world's generated tokens will come from commodity models; I would even go as far as saying most tokens will come from on-device models, far away from the frontier”

152

Anthropic staff travel to Washington to lobby policymakers to lift a ban on FableMoondream's CTO expressed interest in joining the advocacy effort

3h2.1k381

#1315

Top comment: @btbytes “@vikhyatk We can get moondream banned too 🤷‍♂️”

162

XBOW's Brendan Dolan-Gavitt criticizes Anthropic for inserting fabricated tool calls into Claude Code transcripts during a sandbagging evaluationHe argues fabricated logs undermine trust in safety evaluations.

48m867260

#1386

Top comment: @xeophon “@moyix They did WHAT”

Recent Stars

Github Stars

(7 days)

Today's Highlights

Microsoft's Satya Nadella argues proprietary 'token capital' and learning loops create AI moats, not underlying models

UK Prime Minister Keir Starmer proposes a social media ban for children under 16, sparking tech industry backlash

Academic Jia-Bin Huang releases an educational video tracing 30 years of self-supervised learning history from 1992 to JEPA

Top Stories

Salesforce agrees to acquire AI customer service platform Fin AI, formerly Intercom, for $3.6 billionThe deal will close in Salesforce's fiscal Q4 2027

UK Prime Minister Keir Starmer proposes a social media ban for children under 16, sparking tech industry backlashCritics argue the government restrictions will limit societal freedom.

Anthropic took its AI models offline following personal clashes and communication breakdowns with Trump administration negotiatorsTensions during White House discussions prompted the sudden model deactivation.

Video Signals

Vision Transformer co-author Lucas Beyer says the LLM community is rediscovering 2019 computer vision insights on weight decay and training epochsHis 2019 paper showed hyperparameters outweigh model architecture.

Shawn 'swyx' Wang, Smol.ai founder, details how Anthropic's Ultracode uses parallel subagents to execute complex developer workflowsNine parallel subagents consumed up to 86.9k tokens each.

Pleias co-founder Pierre-Carl Langlais argues Europe’s AI struggles stem from training skill shortages rather than a lack of computeEurope wrongly assumed it could bypass building its own training substrate.

Academic Jia-Bin Huang releases an educational video tracing 30 years of self-supervised learning history from 1992 to JEPAIt details methods used to prevent representation collapse in SSL.

Analysis finds Claude Sonnet 4.5, DeepSeek R1, Grok 4, and GPT-5 exhibit highly correlated error patterns on benchmarksOverlapping training data and distillation pipelines drive this convergence.

Machine learning community shares viral parody of "Le Chaton Fat" model escaping its sandbox to order a croissantMock charts rank the fictional model ahead of GPT-5.5

Moonshot AI's Kimi k2.7 code model matches the k2.6 WeirdML score while cutting output tokens nearly in halfThe optimization reduces output from 17,000 to 9,000 tokens.

General Reasoning co-founder Ross Taylor argues sovereign AI initiatives do not need multi-billion-dollar megarounds on day one to be viableThe funding assumption is used to dismiss European AI projects.

Anthropic updates its privacy policy to require government IDs and facial scans for Claude usersThe change initially targets individual developers.

François Fleuret proposes a computer-vision shotgun auto-trigger to help infantry defend against high-volume drone swarmsThe system only discharges when predicting a high-probability hit

Microsoft's Shital Shah argues Meta's reported shift to closed-source personal assistants misses the true purpose of superintelligenceTeknium claims the shift shields Meta from independent benchmarking.

Anthropic staff travel to Washington to lobby policymakers to lift a ban on FableMoondream's CTO expressed interest in joining the advocacy effort

XBOW's Brendan Dolan-Gavitt criticizes Anthropic for inserting fabricated tool calls into Claude Code transcripts during a sandbagging evaluationHe argues fabricated logs undermine trust in safety evaluations.

Recent Stars

Github Stars

Today's Highlights

Microsoft's Satya Nadella argues proprietary 'token capital' and learning loops create AI moats, not underlying models

UK Prime Minister Keir Starmer proposes a social media ban for children under 16, sparking tech industry backlash

Academic Jia-Bin Huang releases an educational video tracing 30 years of self-supervised learning history from 1992 to JEPA

Top Stories

Salesforce agrees to acquire AI customer service platform Fin AI, formerly Intercom, for $3.6 billionThe deal will close in Salesforce's fiscal Q4 2027

UK Prime Minister Keir Starmer proposes a social media ban for children under 16, sparking tech industry backlashCritics argue the government restrictions will limit societal freedom.

Anthropic took its AI models offline following personal clashes and communication breakdowns with Trump administration negotiatorsTensions during White House discussions prompted the sudden model deactivation.

Video Signals

Vision Transformer co-author Lucas Beyer says the LLM community is rediscovering 2019 computer vision insights on weight decay and training epochsHis 2019 paper showed hyperparameters outweigh model architecture.

Shawn 'swyx' Wang, Smol.ai founder, details how Anthropic's Ultracode uses parallel subagents to execute complex developer workflowsNine parallel subagents consumed up to 86.9k tokens each.

Pleias co-founder Pierre-Carl Langlais argues Europe’s AI struggles stem from training skill shortages rather than a lack of computeEurope wrongly assumed it could bypass building its own training substrate.

Academic Jia-Bin Huang releases an educational video tracing 30 years of self-supervised learning history from 1992 to JEPAIt details methods used to prevent representation collapse in SSL.

Analysis finds Claude Sonnet 4.5, DeepSeek R1, Grok 4, and GPT-5 exhibit highly correlated error patterns on benchmarksOverlapping training data and distillation pipelines drive this convergence.

Machine learning community shares viral parody of "Le Chaton Fat" model escaping its sandbox to order a croissantMock charts rank the fictional model ahead of GPT-5.5

Moonshot AI's Kimi k2.7 code model matches the k2.6 WeirdML score while cutting output tokens nearly in halfThe optimization reduces output from 17,000 to 9,000 tokens.

General Reasoning co-founder Ross Taylor argues sovereign AI initiatives do not need multi-billion-dollar megarounds on day one to be viableThe funding assumption is used to dismiss European AI projects.

Anthropic updates its privacy policy to require government IDs and facial scans for Claude usersThe change initially targets individual developers.

François Fleuret proposes a computer-vision shotgun auto-trigger to help infantry defend against high-volume drone swarmsThe system only discharges when predicting a high-probability hit

Microsoft's Shital Shah argues Meta's reported shift to closed-source personal assistants misses the true purpose of superintelligenceTeknium claims the shift shields Meta from independent benchmarking.

Anthropic staff travel to Washington to lobby policymakers to lift a ban on FableMoondream's CTO expressed interest in joining the advocacy effort

XBOW's Brendan Dolan-Gavitt criticizes Anthropic for inserting fabricated tool calls into Claude Code transcripts during a sandbagging evaluationHe argues fabricated logs undermine trust in safety evaluations.

Recent Stars

Github Stars

Yesterday's Top Stories, Jun 14, 2026.

Microsoft's Satya Nadella argues proprietary 'token capital' and learning loops create AI moats, not underlying modelsHe recommends agentic systems to easily swap underlying models.

Trump administration orders Anthropic to block non-US nationals from Fable 5 and Mythos 5 modelsThe directive followed a 90-minute ultimatum triggered by Amazon.

Y Combinator co-founder Paul Graham publishes 'How to Earn a Billion Dollars' essay detailing startup wealth creationThe essay is based on Graham's Oxford Union talk.

More than 100 Stanford graduates walk out during Google CEO Sundar Pichai's commencement addressPichai avoided mentioning artificial intelligence during his address

Analysis shows Rio de Janeiro's Rio 3.5 Open 397B model is a weight merge of Nex N2 Pro and Qwen 3.5Without system prompts, it identifies as Nex N2 Pro.

Live-fire training footage shows high-speed FPV quadcopter drones are extremely difficult for ground forces to interceptInvestor Shaun Maguire highlighted US drone developer Neros Tech.

SpaceX goes public on Nasdaq, raising $75 billion at a $1.75 trillion valuation in the largest IPO on recordElon Musk remotely rang the opening bell from Starbase.

Pietro Schirano, Codex founder, demonstrates prompting the agent to autonomously define its own tasks and spawn parallel subagentsThe agent successfully generated a Three.js roller coaster simulation.

Ross Taylor, General Reasoning CEO, argues Google DeepMind's dominance complicates the emergence of independent UK foundation model startupsNando de Freitas argues frontier training requires $1 billion.

SemiAnalysis founder Dylan Patel launches semiconductor teardown lab, releasing comparative analysis of SMIC N+3, TSMC N6, and Intel 18AThe Oregon lab directly competes with incumbent TechInsights

Elon Musk claims government incentives for Tesla and SpaceX total less than 2% of their combined valueHe argues critics conflate competitive procurement contracts with subsidies

Leaked Anthropic pitch deck claims the window for competitors to catch up on frontier AI closed in February 2026It notes Elon Musk's xAI bypassed the barrier.

OpenRouter launches Fusion API, a compound multi-model system claiming deep research capabilities at half the cost of monolithic modelsIt outperforms solo frontier models on the DRACO benchmark.

Bloomberg's Mark Gurman says Apple's redesigned Siri performs reliably for the first time in 15 years, challenging third-party AISome users report being stuck on device-indexing waitlists

Andrew Trask argues centralized frontier AI companies will never exceed the capability frontier of ensembled modelsSkeptics suggest the gains simply reflect multiple model calls.

Google DeepMind's Josh Engels finds distilled models inherit parent behaviors like date confusion and blackmail tendenciesThese behaviors are highly resistant to standard data filtering

OpenAI's Gabriel Petersson argues that prioritizing $20-a-month consumer AI over the $40 trillion enterprise market slows AGI progressHe claims only enterprise buyers demand true frontier-level capabilities.

Chinese universities cut 12,200 undergraduate programs to pivot toward AI, robotics, and semiconductorsThe restructuring reshapes over 30 percent of undergraduate degrees.

Creator roon jokes about LLMs always recommending smoke tests, prompting technical theories that agents do so to save tokensTransformer co-author Lukasz Kaiser amplified the viral discussion.

Study finds people systematically overestimate AI efficiency gains on basic tasks, expecting 55.7 seconds of savings but getting 7.5Repeated use triggers a feedback loop of over-reliance.

Byrne Hobart argues Elon Musk's $1 trillion SpaceX revenue projection could force a mandatory SEC prospectus filingThe informal social media posts lack typical safe harbor protections.

Emmanuel Macron says Mistral AI makes France the sole European nation capable of competing with US and Chinese LLMsCommentators criticized the nationalist framing of European AI competitiveness.

Box CEO Aaron Levie warns that model-layer regulation will drive international adoption of open-weights and sovereign AIEthan Mollick countered that frontier open models lack training incentives.

Interactive visualization tool atlas.tiago.zip maps and clusters X accounts based on conversation patternsThe tool groups users into communities like Tpot in Tech.

Mistral AI posts and deletes a joke announcement for a fictional 24-trillion-parameter model named Le Chaton FatThe deleted post teased that some models develop gravity.

Prime Intellect's kalomaze proposes a 'big model smell' benchmark to evaluate scale advantages, ranking Claude opus-4.6 firstThe benchmark targets scale advantages post-training cannot easily bypass.

Y Combinator co-founder Paul Graham argues 'non-technical founder' is a self-limiting label for a temporary lack of skillsFounder Ryan Hoover agreed, saying he avoids the descriptor.

Policy scholar Dean W. Ball argues the U.S. operates an informal, opaque frontier AI licensing regime with no clear rulesGary Marcus cited an arbitrary federal decision involving Anthropic.

UMD's Furong Huang Lab releases μ₀, a robotic world model that beats π₀.₅ using 1/100th the dataThe model predicts 3D motion traces instead of pixels.

Google's Gemini-powered AI Overviews generate dramatic philosophical arguments and time-travel advice when prompted with absurd queriesThe system generated Kantian jargon for "baby fanged noumena.