Tech News

10

OpenAI announces Jalapeño, a custom LLM inference chip co-developed with Broadcom that taped out in nine monthsEarly samples are currently running in OpenAI's laboratory.

7h543.2k5.5k490#1COMMENTS429

Top comment: @OpenAI “https://openai.com/index/openai-broadcom-jalapeno-inference-chip/”

20

31

Watchlist

Video Signals

#233

Figma Agent Generates Custom Shaders And Visual Effects

3h ago28K852271

#792

Animator Uses Seedance AI to Render Anime from 3D Previz Preserving Motion

4h ago140K1.7K1K

#80

Lecture Explains ML Foundations Of LLM Post-Training

6h ago17K451531

#888

Gemini 3.5 Flash Adds Computer Use for Cross-Platform Automation

4h ago30K377101

#60

Printing Press Made Written Word Authoritative, Fueling Scientific Revolution

3h ago12K7940

#1247

Medra Introduces AI Experimentalist Agent For Automated Biology Experiments

5h ago12K10232

#766

GPT-Realtime-2 Enables In-Context Wake Words And Simon Says Games

3h ago20K26493

#659

CEO Recommends Award-Winning AI Short Film Costa Verde

6h ago8K148117

#447

Entrepreneur Builds AI Agent Hal That Joins Zoom Meetings

1h ago9.2K6128

#684

DAIR AI Releases Adaptive /Learn Skill For Personalized Learning Hubs

5h ago11K4477

41

Claude Code CLI updates suggest Anthropic is preparing to bundle Fable 5 into standard subscriptionsAnthropic has not yet officially announced the return.

4h307k3.1k255

#682

Top post: @kimmonismus “This Fable 5 update sounds almost too good to be true: The latest Claude Code update does not only include some interesting hints about Fable 5’s return. It also suggests that the model may be permanently added to the subscription plan. That would be fantatic. And I hope it's true insofar as Anthropic generates good PR with it.”

51

New github stars (48hrs)

Rift creates instant copy-on-write workspace clones as a fast alternative to git worktrees

anomalyco/rift6471d

Creates instant copy-on-write clones of directories via btrfs snapshots, reflinks or APFS as git-worktree alternatives.

1014

1325

CEO-Bench evaluates long-horizon LLM agents via 500-day startup simulations with business databases and market dynamics

zlab-princeton/ceobench-src291d

Simulates 500-day startup operations to benchmark long-horizon LLM agents via business databases, tools, and market events.

1014

802

Swift adapter exposing llama.cpp GGUF models through Apple's LanguageModel protocol

pcuenca/LlamaLanguageModels41d

Implements Apple's Foundation Models LanguageModel API for llama.cpp GGUF models in Swift.

1014

861

Collection of normalized two-bar breakbeat audio samples

switchangel/breaks901d

Hosts normalized two-bar breakbeat WAV samples plus Strudel config file.

1014

1398

Lightweight Python library for zero-shot chain-of-thought evaluation of language models on standard benchmarks

openai/simple-evals4.5k1d

Evaluates language models on benchmarks like MMLU, GPQA, MATH, and SimpleQA using zero-shot chain-of-thought prompting.

1014

144

616

70

Arthur Conmy, creator of Automated Circuit Discovery, joins Anthropic to focus on aligning models during trainingHe previously researched mechanistic interpretability at Google DeepMind.

4h144.1k1.7k199

#501

Top post: @livgorton “@ArthurConmy yooo welcome!!”

83

Figma releases Weave Generator, Figma Motion, and an integrated AI agent to build plugins and shaders directly in its canvasThe updates consolidate creative, generative, and code-based canvas workflows

5h120.4k3.6k547

RISING LIKES3.6k

#233

Top post: @zoink “Everything you saw today: code, motion, shaders, your own tools... It all lives together in the canvas. For the first time, in one place. Today we announced ways to: → explore with code → generate with Weave → build plugins and shaders with the Figma agent → animate with Figma Motion”

95

OpenAI study finds reinforcement learning targeting beneficial traits produces robust, emergent alignmentThe alignment gains generalize and resist adversarial pressure.

11h55.4k451186

#178

Top post: @Noahpinion “I continue to suspect that AI, in most cases, will simply be an Average Dude, and that an Average Dude is a Pretty Good Dude”

101

Two recent xAI employees say more than half of Grok usage is driven by adult and NSFW contentThe trend mirrors parent platform X's high NSFW traffic

6h121.8k754128

#57

Top post: @DKThomp “The trillionaire who believes declining fertility is the most important problem in the world has devoted enormous time and treasure toward building what is essentially a next-generation porn website. People are complicated.”

114

John Carmack argues id Software should have upgraded the Doom engine instead of rewriting technology during Quake's developmentHe also admitted pushing the development team to burnout

5h#1VIEWS746.6k#1LIKED8.4k#1BOOKMARKED1.6k

#96

Top comment: @YorchTorchGames “@ID_AA_Carmack I never imagined I’d see Romero (via Doomguy), Sandy, and Carmack here on X reflecting on how Quake shook id Software to its core. Thank you, legends, for one of the greatest games ever made... and the one that defined the 3D FPS genre. Thank you for Quake.”

120

Key Gemini model contributors Jonas Adler and Alexander Pritzel leave Google to join AnthropicThe departures continue a trend of staff leaving Google AI.

2h157.2k951184

#72

Top post: @AndrewCurran_ “I'm quoting this instead of retweeting it because I have things to add to the thread.”

130

Trump administration reportedly prefers meeting with Anthropic co-founder Tom Brown over CEO Dario Amodei for Fable 5 discussionsGovernment satisfaction reportedly improved after Brown led the talks

4h135.4k1.1k102

#57

Top post: @teortaxesTex “one of the two must stay in the locker either Fable, or Dario Dario is a weirdo, so Fable gets to walk”

146

Runlayer raises $30 million from Felicis and Khosla Ventures to secure enterprise AI agents and shadow AIThe platform offers identity-aware controls and runtime security tools

5h125.9k274192

#161

Top post: @vkhosla “This would be an important new category”

154

Google integrates native computer use into Gemini 3.5 Flash with built-in safeguards against prompt injectionThe model scored 78.4 on the OSWorld-Verified benchmark

4h82.4k1.2k373

#684

Top post: @osanseviero “Excited to introduce Computer Use support for Gemini 3.5 Flash!🔥 This enables Gemini to reason and act across platforms (browser, mobile, and desktop environments) We see significant improvements across many work-related automation tasks, from filing tickets and more. Enjoy!”

167

Notion launches External Agents to integrate Claude and Cursor directly into shared team workspacesThe integrated agents can triage bugs and generate pull requests.

4h37.2k45181

#712

Top post: @ryolu_ “use cursor in notion use notion in cursor”

171

OpenThoughts releases OpenThoughts-Agent, open-sourcing a 32B model scoring 26% on Terminal Bench and full training recipesIt features over 100 ablations on RL data curation

3h25.3k189102

#61

Top post: @AlexGDimakis “We are excited to announce what we have been working on for more than six months: The OpenThoughts-Agent dataset and OpenThinker-agent models. More than 100 ablations on data curation for RL environments for coding agents. Our data recipe is SOTA over all open-data agents in their class. We post-train a Qwen-3-32B to get 26% on Terminal Bench and open all our training sets, data pipelines, experiments and models. Some lessons we learned for training agents vs reasoning: 1. The Diversity of tasks matters more, compared to reasoning (OpenThoughts-Agent vs OpenThoughts). You could teach reasoning from math and it transfered widely but RL environments seem to teach more specific capabilities, so each domain must be covered. 2. Filtering high quality and hard questions remains very important. (Was also true for OpenThoughts reasoning). We discuss several ways of filtering. 3. Synthetic re-writing and task augmentation didn’t give significant benefits in our experiments. Sampling multiple teacher rollouts per task did work (was also true for reasoning). Even when keeping the dataset size fixed, multiple answers gave benefits. The Multiple answers mystery is still valid for agentic environments. 4. Stronger models are not necessarily better teachers (was also true for reasoning). The stronger teacher for Quen-3 was GLM-4.7-AWQ and the Terminus2 harness in Daytona. We are releasing 100k tasks and trajectories. 5.Benefits from GRPO remain limited and still on-going. I currently officially hate GRPO.”

182

Nathan Lambert, AI2 post-training lead, releases an ML foundations lecture to kick off his 11-part post-training courseThe video lecture covers KL divergence, cross-entropy, and MDPs.

6h34.5k688650

#80

Top post: @natolambert “Lecture 0: https://www.youtube.com/watch?v=MMDNaeIFVy8&list=PLL1tdVxB1CpVpEtMHxwuR4uI4Lxjw00_y&index=2”

192

Keen Technologies founder John Carmack reflects on how Jerry Pournelle's pessimistic 1970s sci-fi contrasted with the subsequent tech boomSteven Sinofsky highlighted Pournelle's influential Byte magazine column.

7h385.9k2.9k344

#96

Top post: @ID_AA_Carmack “I was delighted the first time he mentioned one of my shareware games in Byte, and it was a pleasure to later get to know him personally when he was a fixture at the Space Access conferences. On SpaceX -- The idea that a single commercial company would eventually be launching over five times as much payload to orbit as every government on earth combined would have seemed outlandish in a story!”

204

Jürgen Schmidhuber, LSTM pioneer, argues Alan Turing's contributions to computer science are oversold compared to Leibniz and GödelResearcher Aran Nayebi countered that Gödel preferred Turing's machines

10h31.2k221111

#86

Top post: @pmddomingos “Schmidhuber is right for once.”

Today's Highlights

Former Greylock partner Josh Elman joins Andreessen Horowitz to focus on consumer startups

John Carmack argues id Software should have upgraded the Doom engine instead of rewriting technology during Quake's development

OpenAI announces Jalapeño, a custom LLM inference chip co-developed with Broadcom that taped out in nine months

Behnam Neyshabur's AI startup Mirendil raises $200M to build self-accelerating systems that automate AI engineering

Top Stories

OpenAI announces Jalapeño, a custom LLM inference chip co-developed with Broadcom that taped out in nine monthsEarly samples are currently running in OpenAI's laboratory.

Stripe and Nan Ransohoff launch Intercept, a $500 million fund to develop broad-spectrum preventatives and air-cleaning technologiesThe initiative will deploy far-UVC light and HEPA filtration.

Qualcomm acquires LLVM and Swift creator Chris Lattner's AI startup Modular, developer of the Mojo languageThe acquisition aims to scale hardware-independent AI compute platforms

Video Signals

Claude Code CLI updates suggest Anthropic is preparing to bundle Fable 5 into standard subscriptionsAnthropic has not yet officially announced the return.

Z.ai's GLM-5.2 scores 22.8% on ARC-AGI-2, leading Chinese models but prompting debate over Western benchmark hill-climbingIt scored 77% on ARC-AGI-1 for $0.19 per task.

Behnam Neyshabur's AI startup Mirendil raises $200M to build self-accelerating systems that automate AI engineeringAndreessen Horowitz and Kleiner Perkins led the seed round

Arthur Conmy, creator of Automated Circuit Discovery, joins Anthropic to focus on aligning models during trainingHe previously researched mechanistic interpretability at Google DeepMind.

Figma releases Weave Generator, Figma Motion, and an integrated AI agent to build plugins and shaders directly in its canvasThe updates consolidate creative, generative, and code-based canvas workflows

OpenAI study finds reinforcement learning targeting beneficial traits produces robust, emergent alignmentThe alignment gains generalize and resist adversarial pressure.

Two recent xAI employees say more than half of Grok usage is driven by adult and NSFW contentThe trend mirrors parent platform X's high NSFW traffic

John Carmack argues id Software should have upgraded the Doom engine instead of rewriting technology during Quake's developmentHe also admitted pushing the development team to burnout

Key Gemini model contributors Jonas Adler and Alexander Pritzel leave Google to join AnthropicThe departures continue a trend of staff leaving Google AI.

Trump administration reportedly prefers meeting with Anthropic co-founder Tom Brown over CEO Dario Amodei for Fable 5 discussionsGovernment satisfaction reportedly improved after Brown led the talks

Runlayer raises $30 million from Felicis and Khosla Ventures to secure enterprise AI agents and shadow AIThe platform offers identity-aware controls and runtime security tools

Google integrates native computer use into Gemini 3.5 Flash with built-in safeguards against prompt injectionThe model scored 78.4 on the OSWorld-Verified benchmark

Notion launches External Agents to integrate Claude and Cursor directly into shared team workspacesThe integrated agents can triage bugs and generate pull requests.

OpenThoughts releases OpenThoughts-Agent, open-sourcing a 32B model scoring 26% on Terminal Bench and full training recipesIt features over 100 ablations on RL data curation

Nathan Lambert, AI2 post-training lead, releases an ML foundations lecture to kick off his 11-part post-training courseThe video lecture covers KL divergence, cross-entropy, and MDPs.

Keen Technologies founder John Carmack reflects on how Jerry Pournelle's pessimistic 1970s sci-fi contrasted with the subsequent tech boomSteven Sinofsky highlighted Pournelle's influential Byte magazine column.

Jürgen Schmidhuber, LSTM pioneer, argues Alan Turing's contributions to computer science are oversold compared to Leibniz and GödelResearcher Aran Nayebi countered that Gödel preferred Turing's machines

Recent Stars

Github Stars

Yesterday's Top Stories, Jun 23, 2026.

Former Greylock partner Josh Elman joins Andreessen Horowitz to focus on consumer startupsElman previously backed Discord and Medium while at Greylock.

Jack Morris leaves PhD to co-found Engram, an AI startup training personalized models that update every minuteThe continuous training approach aims to enable human-like interaction.

Anthropic launches Claude Tag, embedding its AI assistant into Slack channels to act as an active team memberThe integration lets users delegate tasks like writing code.

Claude argues Europeans should install air conditioning, dismissing emissions concerns and cultural objections as outdatedThe post sparked debate linking cooling to GDP growth

Google fires engineer J. Poehnelt, creator of open-source Workspace CLI 'gog', over AI agent concernsThe dismissal preceded Google's own official Workspace CLI announcement.

Meta is developing an experimental prediction markets app called Arena to compete with Polymarket and KalshiThe standalone app currently does not support real-money transactions.

EngramLab teases a new project, drawing praise from SWE-bench creator Ofir Press without releasing technical detailsResearch engineer Winnie Xu also congratulated the team on the launch.

Elon Musk defends DOGE funding reviews, claiming no validated medical funding was halted and legitimate programs went to the State DepartmentCritics of the cuts cited a contested Lancet study.

OpenAI postpones GPT-5.6 release to mid-July as Google DeepMind pauses Gemini 3.5 ProAnthropic has started rolling out Claude Sonnet 5.

Southwest Airlines completes its first commercial flight equipped with Starlink satellite internetThe system delivers continuous gate-to-gate connectivity throughout the flight.

Trump administration pressures Meta, the sole major US holdout, to submit AI models for federal safety reviewsAnthropic was recently ordered to withdraw its Fable model

Palo Alto Networks CEO Nikesh Arora warns AI firms face a structural 'business model trap' driven by high compute and integration costsHigh token prices discourage the experimentation needed for enterprise customization.

Elon Musk shares a viral parody video juxtaposing Representative Alexandria Ocasio-Cortez with a SpaceX Starship Raptor engine testVenture capitalist Sam Lessin also shared the video.

Mistral AI releases Mistral OCR 4, scoring 85.2% on OlmOcrBench for structured document extraction in 170 languagesThe model outputs bounding boxes and inline confidence scores.

Google's Gemini vision model passes an informal benchmark by identifying tiny ticks camouflaged on a poppy seed muffinIt identified insect legs and advised discarding the food.

Seedance 2.5 announced to generate 30-second 4K videos from up to 50 reference inputsCreator Chubby clarified the tool is not yet released.

Baidu open-sources Unlimited-OCR, a 3B-parameter model that parses 40-page PDFs with a constant KV cacheThe model achieved state-of-the-art performance on OmniDocBench.

New York City blocks Waymo autonomous vehicles, drawing comparisons to early regulatory fights over UberCritics argue local taxi services suffer from declining reliability.

Demis Hassabis rejected Facebook's higher acquisition bid after Mark Zuckerberg failed a private AI strategy testZuckerberg equated AI with trends like VR and 3D printing.

Meta launches $299 own-brand AI smart glasses using second-generation Ray-Ban hardwareMeta is also considering cheaper, camera-free audio-only glasses

U.S. Department of Energy commits conditional $17.5 billion to finance 10 Westinghouse AP1000 nuclear reactorsThe initiative aims to meet rising AI power demands.

Zynga founder Mark Pincus releases product playbook advising founders to abandon B+ ideas and prioritize retentionHe developed these scaling strategies while building FarmVille.

A pull request in OpenAI's public Codex repository reveals references to an unreleased GPT-5.6 modelThe code changes prevent the model from becoming the default.

Teknium, Nous Research co-founder, launches a `/learn` command that lets Hermes Agent build and test skills from raw filesIt processes raw directories of code, APIs, and PDFs.

AI2 post-training lead Nathan Lambert releases a free lecture on the evolution of synthetic data and on-policy distillationThe curriculum covers constitutional AI, RLAIF, and rubric-based rewards.

Stanford economist Chad Jones will join the Anthropic Institute to research the economic implications of AIHis leave of absence from Stanford begins June 30.

Hivemind launches an AI-enabled social intelligence platform for group event planning and collective discussionsThe platform is now open for public access.

Five Eyes alliance warns frontier AI models will transform offensive cyber capabilities within monthsSusan Zhang criticized the formal advisory as looking AI-generated

Character.AI's 2021 pitch deck surfaces as a deposition exhibit in Garcia v. Character TechnologiesSutter Hill Ventures partner Kevin Kwok shared the slide

Elon Musk confirms Starmind as the name for a planned SpaceX AI satellite constellationNeither company has released technical specs or deployment timelines