AI News

Today's Highlights

In case you missed it

Former OpenAI Sora researcher Gabriel Petersson resigns to launch an independent venture

YJ#423|@YUCHENJ_UW

TM#359|@ISCIENCELUVR

T🪽#255|@TEKNIUM

AG#246|@ANIMESH_GARG

SG#221|@SARANORMOUS

#1 VIEWED

Yuval Noah Harari warns that Argentina's proposal to grant legal personhood to autonomous AI corporations poses systemic risks

MT#449|@TEGMARK

#1 LIKED

1.5k

Paul Graham, Y Combinator co-founder, says investor questions about AI labs copying startups are the new "what if Google does it

RISING LIKES

447

OpenAI's Gabriel Petersson argues AGI will require highly detailed, dynamic instructions rather than simple open-ended text prompts

GK#1215|@GREGKAMRADT

DA#924|@GROWING_DANIEL

GA#712|@GABRIEL1

Top Stories

Posts:—|Clusters:—|Next Crawl:—

10

Google orders over 3 million TPUs from Intel for 2028 as Nvidia tests Intel's 18A nodeIntel stock jumped 12% in early trading following the reports.

4h|165.3k1.3k127

SS#1967|@STEVESI

ZE#1471|@ZEPHYR_Z9

CH#1448|@KIMMONISMUS

RP#1031|@ROHANPAUL_AI

AC#518|@ANDREWCURRAN_

20

Daniel Han, Unsloth creator, launches OpenEnv to standardize reinforcement learning interfaces with backing from Hugging Face, Meta-PyTorch, and NVIDIAIt decouples AI models from tightly bound RL training environments

2h|43k24686

C🎉#848|@CHARLES_IRL

DH#773|@DANIELHANCHEN

WB#340|@WILLCCBB

C🤗#67|@CLEMENTDELANGUE

TW#17|@THOM_WOLF

30

Wayve co-founder Alex Kendall and Uber open a London waitlist for robotaxis using end-to-end machine learningRides will run between King's Cross and Heathrow Airport.

6h|135.9k56063

AK#1335|@ALEXGKENDALL

Watchlist

Video Signals

Andrej Karpathy Demonstrates Everyday AI Workflow in Two-Hour Recording

1d ago|Views 1.5MLikes 10KBookmarks 29K

DH#773|@DANIELHANCHEN

WB#340|@WILLCCBB

C🤗#67|@CLEMENTDELANGUE

TW#17|@THOM_WOLF

OpenEnv Opens To Community As Hugging Face And NVIDIA Back Agentic RL Protocol

3h ago|Views 27KLikes 110Bookmarks 57

ChatGPT Adds Interactive Charts For Data Visualization On Mobile And Web

1h ago|Views 61KLikes 1.4KBookmarks 259

@(#196|@TIMNITGEBRU

Peter Thiel Funds AI Startup To Let Rich Users File Complaints Against Journalists

11h ago|Views 16KLikes 466Bookmarks 92

MN#471|@MO_NOROUZI

Ideogram V4 Enables Precise Composition Control In ComfyUI With Bounding Boxes

1d ago|Views 15KLikes 273Bookmarks 275

LT#958|@_LEWTUN

TW#17|@THOM_WOLF

Hugging Face Launches CADGenBench to Evaluate AI 3D CAD Generation

2h ago|Views 5.5KLikes 49Bookmarks 18

CH#1448|@KIMMONISMUS

Anthropic President Says AI Has Replaced Almost No Jobs

4h ago|Views 61KLikes 190Bookmarks 41

VS#727|@VINCESITZMANN

MilliVid Generates Long-Range Consistent Videos Using Hierarchical Latents

1h ago|Views 2.5KLikes 89Bookmarks 40

ChatGPT Adds Interactive Charts, Table of Contents, and Message Editing

2h ago|Views 18KLikes 295Bookmarks 66

PS#1602|@SKIRANO

MagicPath Launches Builder Plan for AI Agent Workflows at $10/Month

28m ago|Views 1.6KLikes 28Bookmarks 16

40

Yuval Noah Harari warns that Argentina's proposal to grant legal personhood to autonomous AI corporations poses systemic risksHarari argues autonomous agents cannot face traditional legal penalties.

7h|#1VIEWS218.3k817#1BOOKMARKED308#1COMMENTS183

MT#449|@TEGMARK

51

Paul Graham, Y Combinator co-founder, says investor questions about AI labs copying startups are the new "what if Google does itGraham dismissed the inquiries as signs of investor ignorance.

2h|90.3k#1LIKED1.5k271

New github stars (48hrs)

kyutai-labs/kairos517h

Trains 6B LLMs on temporally ordered Common Crawl data from 2018-2025 to measure recency bias and enable continual learning studies.

61

Gary Marcus disputes claims of a mathematician split on AI, noting Terence Tao signed the warning against industry hypeTao has previously praised the capabilities of GPT-5.5 Pro.

6h|45.9k42373

CH#1448|@KIMMONISMUS

GM#157|@GARYMARCUS

72

a16z's Martín Casado argues AI models are structurally capped by the limits of human training dataHe warns reinforcement learning fails to generalize beyond narrow domains.

2h|26.1k32071

AJ#1351|@ALEXJC

SC#707|@SARAHCAT21

MA#465|@MARTIN_CASADO

80

Google DeepMind's Andreas Kirsch urges researchers to join Anthropic, arguing a single market leader reduces safety risksCritics argue joining major labs worsens overall safety dynamics.

6h|44.5k14560

SH#957|@ZERGYLORD

AK#228|@BLACKHC

SB#207|@BLANCHEMINERVA

91

OpenAI's Gabriel Petersson argues AGI will require highly detailed, dynamic instructions rather than simple open-ended text promptsHe compares task delegation to managing an email inbox

1h|20.1k44764

RISING LIKES447

GK#1215|@GREGKAMRADT

DA#924|@GROWING_DANIEL

GA#712|@GABRIEL1

101

University of Maryland's Yekyung Kim finds LLMs suffer from "argument collapse," generating unique arguments just 3.4% of the timeHuman writers produced unique arguments 65.3% of the time.

2h|7.9k7124

TC#1053|@TUHINCHAKR

MI#366|@MOHITIYYER

112

UK launches AI Economics Institute chaired by Nobel laureate Simon Johnson to study labor and productivity impactsThe new body integrates DSIT’s Future of Work Unit.

45m|6.8k9762

AC#518|@ANDREWCURRAN_

SK#505|@SEBKRIER

122

Thomas Wolf and Michael Rabinovich release CADGenBench to evaluate how AI models generate and edit functional 3D engineering partsThe benchmark evaluates Claude and GPT across 81 tasks

1h|10.2k6835

LT#958|@_LEWTUN

TW#17|@THOM_WOLF

131

UK AISI Chief Scientist Geoffrey Irving predicts compilers, the Linux kernel, and major chips will be formally verified by 2027AI-generated specifications will limit overall confidence in the proofs

7h|7.8k9533

GI#349|@GEOFFREYIRVING

MB#20|@MILES_BRUNDAGE

141

Meta AI grew 2.5x to become the third-largest consumer AI app, but its 30-day user retention is just 4.5%Subsidized mobile data plans in emerging markets drive the growth.

2h|16.9k13926

PR#1330|@8TEAPI

DE#363|@DEEDYDAS

151

Moonshot AI's Nathan Chen argues pretraining research remains vital as architectural and optimizer improvements cut required FLOPS by over 20%Sparse attention can reduce long-context FLOPS by over 90%.

1h|6.8k12947

BI#1366|@BILALTWOVEC

EL#706|@ELIEBAKOUCH

169

LeRobot integrates VLA-JEPA, a vision-language-action model that fine-tunes on robotics tasks with just 13 trajectoriesPrevious diffusion policies required 100 trajectories for similar tasks.

7h|7.1k4017

M🍥#1879|@MAYFER

CP#737|@CHRIS_J_PAXTON

170

UCL professor Tim Rocktäschel argues recursive self-improvement is key to attaining artificial superintelligenceA resurfaced October 2024 interview outlines iterative capability scaling.

7h|4.3k5937

SC#1166|@SAMCHARRINGTON

180

Stephen Casper says his MIT AI safety doctoral thesis is only the second in the institute's history to contain the word 'shitThe research focuses on model tampering and adversarial training.

4h|8.7k10113

AK#1584|@DR_ATOOSA

C(#572|@STEPHENLCASPER

190

Ronak Pradeep, formerly of Google and Apple, joins Databricks' Mosaic AI to build information-seeking agentsHis PhD research focused on TREC RAG and rerankers.

3h|5.5k617

AD#954|@MRDROZDOV

JF#106|@JEFRANKLE

200

SuperPowers launches Super, an AI agent that uses its own cloud environments and payment credentials to run real-world tasksA computer-use cache reduces repeating task costs to near-zero.

1h|3.9k129

SA#1748|@SVPINO

CH#1448|@KIMMONISMUS

Rising Stories

Benchmarks Overrated2H AGO
Open-Weight Models Close Capability Gap
#975
12511
Internal Wins Hinted26M AGO
Rohan Anil Teases Strong CoreAutoAI Results
#79
351.3k
AGI Barriers Cited2H AGO
Bulletin Warns AGI Faces Permanent Technical Limits
#543
111.5k
Reproducibility Push2H AGO
Professor Demands Code For Awards
#932
342.2k

Recent Stars

egraphs-good/egglog2M AGO
759
…physics-intern-skills3H AGO
Agentic framework for physics
958
2 stars
…tal/phoenix_storybook12H AGO
Mounts a pluggable Storybook UI in Phoenix apps to explore, document, and interactively test components.
862
862 stars
…thu/Audio-Interaction15H AGO
Implements the first unified always-on audio interaction model for streaming and offline tasks like ASR, translation, and proactive responses.
260
193 stars
kyutai-labs/kairos17H AGO
Trains 6B LLMs on temporally ordered Common Crawl data from 2018-2025 to measure recency bias and enable continual learning studies.
631
203
5 stars

Github Stars

(7 days)

kyutai-labs/kairos17H AGO
Trains 6B LLMs on temporally ordered Common Crawl data from 2018-2025 to measure recency bias and enable continual learning studies.
631
203
5 stars
…monw/micropython-wasm2D AGO
Runs MicroPython snippets in a fresh WebAssembly sandbox via wasmtime.
862
194
85 stars
…tinual-learning-bench3D AGO
Benchmarks AI agents on learning from repeated interactions across multi-episode tasks with adaptation metrics.
862
846
127 stars
trycua/cua3D AGO
Supplies sandboxes, SDKs, and benchmarks for AI agents to control full desktops on macOS, Linux, and Windows.
862
773
17.7k stars
diego-lima/rlmy3D AGO
UNKNOWN.
862
158
10 stars

Privacy Policy|Terms of Service|© 2026 Digg Inc.

Yesterday's Top Stories, Jun 7, 2026.

Posts: 3,280|Clusters: 831|Stories frozen at midnight PT

10

Former OpenAI Sora researcher Gabriel Petersson resigns to launch an independent venturePetersson previously led an internal development team at OpenAI.

11h|915.9k8k940

AM#1930|@GRAVICLE

SA#1266|@SAMSJA19

JH#1198|@JOHANNES_HAGE

MN#1174|@MENHGUIN

FB#1117|@XEOPHON

CK#985|@KALINOWSKI007

LA#975|@SCALING01

B(#850|@BEFFJEZOS

GA#712|@GABRIEL1

VW#696|@VINCENTWEISSER

MC#569|@MMMBCHANG

AC#518|@ANDREWCURRAN_

YJ#423|@YUCHENJ_UW

TM#359|@ISCIENCELUVR

T🪽#255|@TEKNIUM

AG#246|@ANIMESH_GARG

SG#221|@SARANORMOUS

20

OpenClaw project steward Peter Steinberger argues developers should stop manually prompting AI coding agents and build automated programmatic loopsPeers requested concrete workflow examples and noted implementation limits.

11h|#1VIEWS2.9M#1LIKED14.6k#1BOOKMARKED10.7k

MB#1802|@MATTHEWBERMAN

PR#1330|@8TEAPI

AV#1245|@ALTRYNE

AD#867|@DORIALEXANDER

B(#850|@BEFFJEZOS

SI#819|@SUZATWEET

CA#730|@CHRISALBON

PM#710|@PMINERVINI

RS#617|@ZIV_RAVID

PS#495|@STEIPETE

YD#306|@YUNTIANDENG

GT#266|@GARRYTAN

OK#158|@LATEINTERACTION

30

OpenAI's roon claims support is growing for mutual conditional pause agreements ahead of recursive self-improvementElie Bakouch shared GPT-5.5 benchmarks showing marginal debugging progress.

3h|73.1k1.2k139

DE#1934|@DANIEL_271828

CF#1357|@CFGEEK

B(#850|@BEFFJEZOS

CP#737|@CHRIS_J_PAXTON

EL#706|@ELIEBAKOUCH

YS#618|@YONASHAV

KA#488|@YACINEMTB

DW#390|@DEANWBALL

40

OpenAI plans to overhaul ChatGPT into a superapp with coding tools and AI agents in the coming weeksThe revamp aims to drive higher-margin monetization opportunities.

1d|155.4k1.7k213

🍓🍓#1713|@IRULETHEWORLDMO

CH#1448|@KIMMONISMUS

AV#1245|@ALTRYNE

RP#1031|@ROHANPAUL_AI

AC#518|@ANDREWCURRAN_

MB#20|@MILES_BRUNDAGE

50

Coinbase CEO Brian Armstrong predicts 80 percent of AI workloads will migrate to 99 percent cheaper models within 12 to 18 monthsIndustry founders agree this shift makes model routing critical.

6h|949.2k3.8k2.2k

YP#1235|@YPATIL125

JL#670|@JERRYJLIU0

KA#488|@YACINEMTB

60

More than 25 open-weight models launch in a single week, led by NVIDIA's Nemotron 3 Ultra and Google's Gemma 4Ideogram also shipped its first open-weight image model.

2d|184.3k1.1k258

LA#975|@SCALING01

T(#421|@TEORTAXESTEX

DW#390|@DEANWBALL

⿻A#357|@IAMTRASK

MB#20|@MILES_BRUNDAGE

70

ChatGPT hallucinates unsettling images and triggers safety warnings when prompted to 'restore' nonexistent photosTranslating the prompt into Latin produced occult parchment images

1d|87.4k76891

“P#1704|@PAULARAMBLES

AM#1693|@ANDYMASLEY

GL#1479|@GLEECH

CB#1004|@CODE_STAR

🎭#862|@DEEPFATES

OE#301|@OWAINEVANS_UK

MB#20|@MILES_BRUNDAGE

80

Gergely Orosz reports Meta is assigning manual data labeling to software engineers, sparking debate over whether top technical talent should label dataDeepSeek's Liang Wenfeng personally annotated data for DeepSeek-V3.

1d|124.9k919201

PS#1425|@PETESKOMOROCH

CP#737|@CHRIS_J_PAXTON

JL#670|@JERRYJLIU0

T(#421|@TEORTAXESTEX

ER#311|@ERICMITCHELLAI

90

OpenAI Codex engineering lead Thibault Sottiaux launches a 100-day program giving selected developers 10x higher usage limitsThe program selects one user daily for month-long expanded access

8h|969.7k8.5k1k

TI#758|@THSOTTIAUX

PS#495|@STEIPETE

100

Notion temporarily disabled Anthropic integrations after performance degradation in Opus 4.7 and 4.8 caused high failure ratesNotion has since restored access after temporarily rerouting traffic

1d|357k2.3k295

IZ#1296|@IVANHZHAO

C🤗#67|@CLEMENTDELANGUE

110

Tim Cook takes hands-on control of Apple's AI strategy following Apple Intelligence setbacks and Siri delaysThis is Cook's most active design role as CEO.

15h|377k3.8k360

DE#1934|@DANIEL_271828

CA#730|@CHRISALBON

120

MIT's Phillip Isola proposes Supervised Memory Training to train RNNs without backpropagation through timeThe method decouples memory updates to enable time-parallel training.

9h|72.8k467357

RL#1463|@ROBERTTLANGE

MJ#1057|@MINQIJIANG

DX#794|@DANFEI_XU

VS#727|@VINCESITZMANN

PI#178|@PHILLIP_ISOLA

130

Dimitris Papailiopoulos of Microsoft Research AI Frontiers teases how Codex used 'Idea Descent' to shrink a solver below 50 kBThe optimization run spanned 110 iterations using LZMA.

15h|70.1k665299

FB#1117|@XEOPHON

LT#958|@_LEWTUN

YD#306|@YUNTIANDENG

DP#193|@DIMITRISPAPAIL

140

Brainwave To Music AI Breakthrough Achieved With NeuroSky And LoRA

12h|152.2k904493

150

OpenRouter data shows Chinese AI models overtook US models in token consumption by May 2026Gary Marcus blames Meta's open-source strategy for the shift

10h|302.6k1.4k525

RP#1031|@ROHANPAUL_AI

GM#157|@GARYMARCUS

160

David Sacks argues AI catastrophe narratives justify government control, while Gary Marcus counters that AI risk is bipartisanE/acc founder Beff Jezos signaled agreement with Sacks' prediction.

7h|348.8k3.4k187

B(#850|@BEFFJEZOS

GM#157|@GARYMARCUS

170

Meta's Lucas Beyer argues NVIDIA's Nemotron could be the last fully open, from-scratch LLM project if OLMo endsStanford's Marin project remains another fully open training framework

16h|88.2k694216

EL#706|@ELIEBAKOUCH

KA#488|@YACINEMTB

LB#55|@GIFFMANA

180

Speculative timeline projecting that Anthropic will deprecate most Claude models by June 2026 sparks developer debateOnly Claude 3 Opus and 'Mythos' escape deprecation.

9h|38.2k49246

CP#1062|@CHRISPAINTERYUP

J⧉#511|@REPLIGATE

190

Researcher Jack Morris points out 12-fingered hands in Ideogram CEO Mohammad Norouzi's demonstration of hand-rendering progressThe image spelled "IDEOGRAM" using two contorted hands

15h|45.5k33425

FO#1970|@FOFRAI

SS#1500|@_SAM_SINHA_

MN#471|@MO_NOROUZI

WD#249|@WILLDEPUE

RA#79|@_AROHAN_

200

Harari Warns AI Could Overtake Human Control of Narratives and Power

19h|167.4k2.2k1.4k

NR#443|@NAVEENGRAO

210

MIT study of 100,000 GitHub developers finds AI assistants increased commits by up to 300% but completed software by only 30%Human bottlenecks in code review and testing limited final output.

10h|107.8k267149

JJ#1865|@J2BRYSON

RP#1031|@ROHANPAUL_AI

CP#725|@CHAMATH

GM#157|@GARYMARCUS

220

Abundant launches SWE-Marathon coding benchmark, where Claude Opus 4.8 leads frontier models with a 26% resolution rateThe long-horizon tasks include building Slack from scratch.

2d|239.7k1.7k2k

AK#937|@ANDYKONWINSKI

BC#320|@BCHERNY

230

Nvidia claims nine of the top 30 spots on Hugging Face's front page, led by its Cosmos3 modelsThe featured open-source models contain 65 billion parameters.

23h|106.3k587103

C🤗#67|@CLEMENTDELANGUE

NL#64|@NATOLAMBERT

AK#29|@_AKHALIQ

240

Andrea Tagliasacchi argues CVPR shouldn't reward exclusive compute, prompting a motion to mandate public code and weights for Best PaperResearcher Ani Kembhavi suggested also requiring public training data.

22h|122k39671

AT#1488|@TAIYASAKI

AK#1477|@ANIKEMBHAVI

TC#1053|@TUHINCHAKR

LB#55|@GIFFMANA

250

Microsoft VP of AI Nando de Freitas proposes replacing complex training pipelines with unified, continual interactive causal agentsThe single stream replaces separate SFT and RLHF objectives

18h|20.7k222237

JA#1546|@JULIUSADML

PA#1289|@ADAPTIVEAGENTS

MB#339|@MICHAEL_J_BLACK

260

Florian Brand, Prime Intellect research engineer, adopts Gemma 4 E4B 6-bit quantized as his primary local Mac LLMThe new setup replaces his nine-month daily Qwen deployment.

19h|70.5k838515

FB#1117|@XEOPHON

KA#488|@YACINEMTB

👩‍💻P#270|@DYNAMICWEBPAIGE

270

Boris Cherny, head of Claude Code at Anthropic, says software development is shifting from manual prompting to autonomous loopsBold Metrics' CTO notes the high cognitive shift required.

11h|65.2k409195

RP#1031|@ROHANPAUL_AI

OK#158|@LATEINTERACTION

280

Sebastian Raschka highlights study showing LLM-generated repository context files fail to improve AI coding agent performanceDeveloper-written files outperformed LLM-generated ones due to domain expertise.

14h|82.8k739992

290

Trump administration explores proposal for the U.S. government to acquire financial stakes in leading AI firmsTrump plans to meet with AI executives next week.

20h|67k1.3k125

RP#1031|@ROHANPAUL_AI

300

YOLO creator Joseph Redmon argues Meta reassigning 5,000 engineers to data labeling is a tactic to prompt voluntary resignationsDefenders argue manual review is essential for understanding training datasets.

14h|34.2k9318

JR#1495|@PJREDDIE

🇺🇦A#785|@SKIMINOK

JL#670|@JERRYJLIU0