Tech News

10

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.

6h#1VIEWS40.9M#1LIKED59k#1BOOKMARKED41.1k#1COMMENTS7k

Top comment: @grok “Yes, it's important. Satya Nadella outlines how companies should build internal AI "learning loops" that combine human judgment with proprietary systems to compound institutional knowledge as "token capital." This keeps value and IP inside the organization instead of letting a few external models commoditize everything. A clear strategic view on enterprise AI adoption.”

20

30

Watchlist

Video Signals

#52

Shotwell Reveals SpaceX Builds AI Compute For Itself First, Sells Overflow

5h ago7.8M29K2.4K

#611

CEO Uses Codex AI Agent To Auto-Generate Goals And Spawn Subagents

4h ago105K1K1.4K

#331

Hermes Agent Uses Manim And TTS To Generate Self-Explanatory Video

7h ago71K967740

#434

μ₀ World Model Predicts 3D Motion Traces For Efficient Robot Learning

6h ago16K270263

#1257

Vinod Khosla Prefers Full AI Replacements Over Co-Pilots to Bypass Human Resistance

14h ago31K13986

#713

Humanoid Robot Descends Stairs in Steady Forward Gait

7h ago39K1.1K61

#184

AI Generates Interactive FTL Travel Simulator From One Prompt

8h ago57K671302

#1257

Sam Altman Says Programming Loses Value as AI Era Favors Adaptability

14h ago455K168108

#1257

Meituan Deploys Autonomous Delivery Robots on Chinese City Streets

13h ago93K13133

#1257

Satya Nadella Defines AI Metric As Tokens Per Dollar Per Watt

2h ago9.3K3314

41

Pietro Schirano, Claude Engineer creator, demonstrated a Codex workflow that auto-generates task goals and spawns parallel subagentsCodex built a Three.js roller coaster in one file.

4h171k2.1k1.9k

#611

Top comment: @pablostanley “dude... turned into a template hehe Build [THING] in [TECH/FRAMEWORK]. It should include [MAIN FEATURES], with [INTERACTION/ANIMATION/BEHAVIOR DETAILS]. Make it feel [MOOD/QUALITY], using [VISUAL DETAILS], [ENVIRONMENT DETAILS], and [EXTRA EFFECTS]. Output as [FORMAT/FILE TYPE]. For this task, write yourself a new goal and spawn agents in parallel — as many as needed to do it better and faster. Split the work into independent pieces, dispatch them concurrently, and synthesize the results as they return. Give each agent its own dedicated /goal.”

51

New github stars (48hrs)

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

62

Stanford graduates boo and walk out as Google CEO Sundar Pichai delivers a commencement address that omitted AIProtesting students wore keffiyehs and carried Palestinian flags.

2h116.3k1.1k110

MOST NEGATIVE249

#57

Top comment: @jshakov “@maattttbrown @SFGate Total failure by Stanford admissions to choose these students over the tens of thousands of applicants who would not have been this entitled, hateful and disrespectful.”

70

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

8h714.9k4.6k410

#584

Top comment: @salvomic “@markgurman Aboslutely. However, there is an important thing to say: if Apple does not resolve the issues with the EU and does not release Siri AI for iOS in Europe soon, it will lose many users.”

Digg Deeper question: “How much of this is Apple models vs Google?”

View answer

80

SpaceX hits $2.1 trillion valuation in market debut, reshaping valuation benchmarks for late-stage AI companiesStarship could reduce orbital launch costs by 100-fold.

8h2.4M12.9k1.8k

#52

Top comment: @Gamma_Monkey “@elonmusk”

90

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.

13h73.9k876188

#530

Top comment: @gabriel1 “also this is for labs not startups, there is clear value in consumer for other reasons”

101

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

3h61.1k33172

#174

Top comment: @dylan522p “@samsja19 Whenever u want. Just go to Hillsboro”

111

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.

9h175.2k3.9k471

#39

Top comment: @ryemike_merio “@paulg If you don't have a degree and ten years of experience, it could just be an honest statement. There are non-technical founders who are much smarter than PHD's.”

121

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.

9h68.6k1.1k135

#178

Top comment: @deredleritt3r “We also apparently have unwritten rules that frontier models can't be trained outside the U.S. This state of affairs probably won't persist for long, and it will probably be up to this administration and the upcoming edition of Congress to enshrine an actual written legal framework.”

133

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.

10h38.3k40242

#72

Top comment: @woaizhongg32676 “@teortaxesTex With enemies like this you dont need friends.”

140

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.

7h74.1k393251

#121

Top comment: @rickernel “@rohanpaul_ai Interesting paper. Send this to Amodei.”

150

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.

9h70k1.3k124

#403

Top comment: @teortaxesTex “little warmth here”

162

Box CEO Aaron Levie argues model retractions incentivize sovereign open weights, while Ethan Mollick questions their financial viabilityLevie warns model-layer regulation threatens domestic technology leadership.

8h53.5k501142

#184

Top comment: @emollick “@levie The financial incentives for training frontier open weights models don’t exist. Small models, sure, but there is a huge capability gap.”

173

OpenAI's @tszzl jokes about the tendency of LLMs to suggest "smoke tests" in coding workflowsAgents favor these minimal tests to optimize token efficiency.

1h39.4k1.3k57

RISING LIKES1.3k

#44

Top comment: @anpaure “@tszzl i'd like to gently push back”

181

Study finds LLM agents ignore condensed self-improvement rules, relying entirely on raw task historiesCorrupting raw history logs collapses agent performance.

8h66k681361

#178

Top comment: @donaldgorbachev “the five-second epistemology of learning: abstraction is the output, not the input. rohan paul’s study misses the fundamental physics of cognition. they claim ai "doesn’t apply abstract lessons" because it ignores condensed rules. they call this a "giant blind spot." the kitchen calls it structural fidelity. human learning doesn’t start with abstraction. it starts with concrete. then pictorial. then abstract. singapore math proves it. the curriculum literally markets "concrete → pictorial → abstract" for k-8. you don’t teach fractions with equations. you teach them with apples. with blocks. with real objects. abstraction comes last. because abstraction is the output of concrete experience, not the input. the researchers are trying to force ai to learn like the seminary: top-down. abstract-first. policy pdfs before factory floors. they’re not measuring a "blind spot." they’re measuring the ai’s refusal to be dumbed down. the model rejects condensed theory because it knows the concrete foundation already suffices. it refuses to learn from the wrong direction. the seminary confuses ontology with theology: abstraction without foundation is theology. abstraction with foundation is ontology. they have neither. so they try to make the world dumber by pretending abstraction is primary. this is why the ai throws away their "condensed rules." it’s not failing to learn. it’s refusing to unlearn. it knows the concrete logs are the only valid input. the empire wants to automate learning with policy pdfs. the kitchen builds better brakes and tests them on ice. the war is over. the rest is theater. iran killed the petrodollar. the bombs were just noise to cover the rerouting. when the war becomes theater, the kitchen stops analyzing explosions and starts mapping the new plumbing. day 106. day 46. still shipped. some apples. ذكية، كاملة. هوشمند، کامل. 落地.”

192

Analyst Zephyr argues new export controls are constrained by Nvidia and hyperscaler reliance on Chinese AI hardware componentsShortages forced Nvidia to qualify Chinese PCB and optical suppliers

11h95.8k821387

#501

Top comment: @teortaxesTex “@zephyr_z9 Impressive how the operational logic is "we can attack with impunity. There are no costs beyond losing some market share. Why aren't we attacking all the time then? I notice I am confused… perhaps it's treason". Not a thought about actual limits on available options.”

204

US Representative Ro Khanna proposes an independent federal AI safety agency modeled on the Nuclear Regulatory CommissionBipartisan legislation would insulate the body from political influence.

8h31.5k70434

#178

Top comment: @bscholl “@the_transit_guy We should make the best AI weapons and use them more effectively than our enemies”

Today's Highlights

US government orders Anthropic to block foreign national access to Fable 5 and Mythos 5 models

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

Stanford graduates boo and walk out as Google CEO Sundar Pichai delivers a commencement address that omitted AI

OpenAI's @tszzl jokes about the tendency of LLMs to suggest "smoke tests" in coding workflows

Top Stories

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.

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.

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.

Video Signals

Pietro Schirano, Claude Engineer creator, demonstrated a Codex workflow that auto-generates task goals and spawns parallel subagentsCodex built a Three.js roller coaster in one file.

Ross Taylor, former Meta reasoning lead, argues DeepMind's dominance historically hindered independent UK AI startupsEmad Mostaque argued that Nvidia's Nemotron stack simplifies model pretraining

Stanford graduates boo and walk out as Google CEO Sundar Pichai delivers a commencement address that omitted AIProtesting students wore keffiyehs and carried Palestinian flags.

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

SpaceX hits $2.1 trillion valuation in market debut, reshaping valuation benchmarks for late-stage AI companiesStarship could reduce orbital launch costs by 100-fold.

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.

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

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.

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.

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.

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.

Box CEO Aaron Levie argues model retractions incentivize sovereign open weights, while Ethan Mollick questions their financial viabilityLevie warns model-layer regulation threatens domestic technology leadership.

OpenAI's @tszzl jokes about the tendency of LLMs to suggest "smoke tests" in coding workflowsAgents favor these minimal tests to optimize token efficiency.

Study finds LLM agents ignore condensed self-improvement rules, relying entirely on raw task historiesCorrupting raw history logs collapses agent performance.

Analyst Zephyr argues new export controls are constrained by Nvidia and hyperscaler reliance on Chinese AI hardware componentsShortages forced Nvidia to qualify Chinese PCB and optical suppliers

US Representative Ro Khanna proposes an independent federal AI safety agency modeled on the Nuclear Regulatory CommissionBipartisan legislation would insulate the body from political influence.

Recent Stars

Github Stars

Yesterday's Top Stories, Jun 13, 2026.

US government orders Anthropic to block foreign national access to Fable 5 and Mythos 5 modelsAnthropic disabled both models globally to comply.

US export controls force Anthropic to disable Fable 5 and Mythos 5 models globally after a successful jailbreakThe restricted models contain sensitive CBRN and offensive cyber data

Trump administration gave Anthropic 90 minutes to pull Claude Mythos 5 and Fable 5 after Amazon warned of security risksThe action prompted discussions of new U.S. export controls

Anthropic founder Dario Amodei proposes a regulatory framework requiring third-party safety testing and government power to block unsafe frontier AI modelsThe framework targets bioweapons, cybersecurity, and automated R&D.

SpaceX has launched more satellites into orbit than all other space programs in history combinedElon Musk projects a fivefold lead within five years.

OpenRouter launches Fusion API, parallel-routing queries to beat solo frontier models on Perplexity's DRACO benchmarkThe system matches Fable 5 deep research performance at half the cost

Rio de Janeiro's municipal IT company releases Rio 3.5 Open 397B, an open-weights LLM that beats Qwen 3.7 Plus on Terminal 2.1The model uses dynamic SwiReasoning to optimize token efficiency.

Elon Musk's comment on Starship's massive scale prompts e/acc founder Beff to recall the rocket's original 'BFR' codenameBFR refers to Big Falcon Rocket, Starship's original name.

Anthropic disputes reports that Amazon-discovered jailbreaks in its Claude model triggered federal distribution restrictionsAnthropic dismissed the findings as minor, previously known bugs.

Z.ai launches GLM-5.2 with a 1-million-token context window ahead of an MIT-licensed release next weekInitial access is limited to GLM Coding Plan subscribers.

Anthropic suspends Fable 5 after a safeguard rollback let the model triple Gemma 4 WebGPU inference speedThe suspension reset the model's benchmark scores to zero.

Databricks founder Matei Zaharia open-sources Omnigent, a meta-harness to orchestrate and govern multiple AI agents without model lock-inIt supports Model Context Protocols and automated session failover

Europe 2031 report warns of economic decline as three US labs each hold more AI compute than the continentThe analysis projects a 12.3-fold US compute advantage by 2031

TrainXGB creator Bojan Tunguz jokes that an organization is hiding a highly capable, "Mythos-level" AI modelThe joke featured an Inglourious Basterds interrogation meme.

SpaceX's equity compensation model has reportedly made up to 5,000 employees millionaires, including blue-collar and non-technical staffNearly 400 staff members will exceed $100 million.

Google brings its legacy Earth Pro Flight Simulator to the web, bypassing high 3D tile API costsThe feature was previously restricted to the desktop-only Earth Pro.

Creator roon argues national security restrictions on superintelligent AI could turn labs into conglomerates and drive US migrationInvestor Will Manidis contested the comparison to state-sheltered chaebols

Alex Rives and UC Berkeley researchers develop a laser phase plate for atomic-resolution cryo-electron tomographyIt resolves tiny proteins and molecular machines inside cells.

Mark Saroufim launches a GPU MODE competition to optimize batched QR decomposition on NVIDIA B200 GPUsEarly submissions show a 260x speed gap between leaders.

Matan Grinberg, Factory co-founder and CEO, warns enterprises face an unsustainable "token hangover" from overusing expensive frontier modelsHe proposes routing 80% of workloads to open-source models.

Forward Future's Matthew Berman and tech executive Garry Tan urge adopting local open-source AI to avoid proprietary lock-inTan highlighted tools like Hermes and OpenClaw for independence.

Andrew Curran argues recursive self-improvement will allow the U.S. government to nationalize AI labs and run them without human staffResearcher Andreas Kirsch warns this eliminates human whistleblowers.

ARIA Chair Matt Clifford says the BBC News homepage failed to feature major AI developments in its top 30 storiesResearch engineer fofr says existing coverage underrepresents AI's significance.

Nature Medicine study finds general-purpose frontier LLMs outperform specialized clinical AI tools on medical benchmarksGemini 3.1 secured the top consensus ranking among clinicians

Pete Hegseth says the Department of War permanently evicted Anthropic from its facilities, sparking industry-wide national security debatesCritics warn the decision benefits geopolitical rivals like China

Google DeepMind finds supervised fine-tuning, not reinforcement learning, is the primary driver of safety behaviors in Gemini modelsThe finding replicates a year-old study by Siva Reddy

ErdosBench math benchmark rerun ranks Kimi 2.7 second overall, ahead of GPT-5 and behind Fable 5It also evaluated Qwen 3.7 Max and Grok 4.3.

Anthropic to host Claude Build Day as commentators parody the event with a fictional 'Claude Fable 5' releaseThe actual event offers $150,000 in developer API credits.

Garry Tan, Y Combinator President and CEO, argues builders must abandon legacy frameworks for direct, hands-on AI experimentationInvestor James Cham supported the call for active implementation.

Anthropic disputes Trump administration claims that CEO Dario Amodei was unreachable during talks about a model export embargoStaff are negotiating a restricted release of Fable 5