AI News

Today's Highlights

In case you missed it

Google releases Gemma 4 12B, an encoder-free multimodal model under Apache 2.0 that beats the larger Gemma 3 27B

PT#250|@ELDER_PLINIUS

SP#10|@SUNDARPICHAI

DH#4|@DEMISHASSABIS

#1 VIEWED

NVIDIA releases Nemotron 3 Ultra, a 550-billion parameter open-weight hybrid Mamba2-Transformer MoE model

P(#461|@RAJAMMANABROLU

C🤗#68|@CLEMENTDELANGUE

NL#64|@NATOLAMBERT

AK#28|@_AKHALIQ

RISING LIKES

257

AI2's Nathan Lambert says Nvidia's multi-teacher on-policy distillation for Nemotron 3 Ultra is the post-training industry standard

WH#1430|@NREHIEW_

FI#397|@FINBARRTIMBERS

YY#128|@YISONGYUE

NL#64|@NATOLAMBERT

It’s a new day at Digg. Fresh stories are clustering now. 85% complete.

Top Stories

Posts:—|Clusters:—|Next Crawl:—

10

NVIDIA releases Nemotron 3 Ultra, a 550-billion parameter open-weight hybrid Mamba2-Transformer MoE modelIt features 55 billion active parameters and multi-token prediction.

2h|#1VIEWS186k#1LIKED2.8k#1BOOKMARKED682#1COMMENTS293

WH#1430|@NREHIEW_

AV#1245|@ALTRYNE

FB#1153|@XEOPHON

VW#707|@VINCENTWEISSER

ME#675|@MERVENOYANN

RS#612|@ZIV_RAVID

YO#519|@Y0B1BYTE

KA#488|@YACINEMTB

EL#475|@OMARSAR0

P(#461|@RAJAMMANABROLU

C🤗#68|@CLEMENTDELANGUE

NL#64|@NATOLAMBERT

AK#28|@_AKHALIQ

20

Ramp raises $750 million at a $44 billion valuation and launches tools to monitor enterprise AI token consumptionCEO Eric Glyman calls tokens the fastest-growing business cost.

1h|117.4k948307

PO#1862|@PATRICK_OSHAG

FB#1153|@XEOPHON

VW#707|@VINCENTWEISSER

30

AI2's Nathan Lambert says Nvidia's multi-teacher on-policy distillation for Nemotron 3 Ultra is the post-training industry standardThe pipeline uses over 10 specialized teacher models.

1h|16.5k257155

RISING LIKES257

WH#1430|@NREHIEW_

FI#397|@FINBARRTIMBERS

YY#128|@YISONGYUE

NL#64|@NATOLAMBERT

Watchlist

Video Signals

RS#617|@ZIV_RAVID

GM#157|@GARYMARCUS

Viral Video Mocks Fragile AI Agent Workflows With Rube Goldberg Machine

1d ago|Views 1.9MLikes 22KBookmarks 4.3K

CH#1448|@KIMMONISMUS

WH#1409|@NREHIEW_

VW#696|@VINCENTWEISSER

KA#488|@YACINEMTB

EL#483|@OMARSAR0

NVIDIA Ships Nemotron 3 Ultra 550B MoE Open Model for Agents

2h ago|Views 83KLikes 1.1KBookmarks 281

EM#74|@ELONMUSK

SpaceX Shares IPO Roadshow Presentation Highlighting Starship And AI Ventures

43m ago|Views 788KLikes 6.8KBookmarks 1.1K

CH#1448|@KIMMONISMUS

EL#483|@OMARSAR0

RS#323|@SCOBLEIZER

Engramme Unveils Large Memory Models to Solve AI's Persistent Memory Problem

Apr 28|Views 1.1MLikes 1.5KBookmarks 1.8K

CM#16|@CHRMANNING

Gemma 4 12B Runs Locally on 8GB RAM via Unsloth Dynamic GGUFs

5h ago|Views 65KLikes 841Bookmarks 528

AJ#1351|@ALEXJC

Interactive Visualization Demonstrates Principal Component Analysis Optimization

6h ago|Views 136KLikes 1.1KBookmarks 657

MN#1174|@MENHGUIN

Sichuan Dancer Earns AGT Pass With Eight Unitree Humanoid Robots

1d ago|Views 78KLikes 636Bookmarks 169

AS#120|@ARAVSRINIVAS

Perplexity Computer Connects Businesses To 400+ Tools For Automation

22h ago|Views 36KLikes 398Bookmarks 80

T🪽#255|@TEKNIUM

Nous Research Joins Nvidia Nemotron Coalition To Advance Open Models

21m ago|Views 6.4KLikes 306Bookmarks 33

RS#617|@ZIV_RAVID

Viral Video Satirizes Companies' Clunky AI Agent Workflows

2h ago|Views 1.7MLikes 0Bookmarks 0

40

Hugging Face co-founder Julien Chaumond releases SynthTraces, a minimal codebase for generating synthetic coding agent session tracesThe tool uses llama.cpp to simulate human-agent interactions.

2h|10.8k13070

SI#825|@SUZATWEET

JC#251|@JULIEN_C

C🤗#68|@CLEMENTDELANGUE

50

Policy scholar Dean W. Ball argues that fully autonomous corporations may require a global banThe debate was sparked by Argentina's national AI strategy.

1h|5.9k848

SK#505|@SEBKRIER

DW#392|@DEANWBALL

DD#53|@DAVIDDUVENAUD

New github stars (48hrs)

Omnimodal world models unifying reasoning and multimodal generation for physical AI

NVIDIA/cosmos8.7k1d

Hosts omnimodal Mixture-of-Transformers models for joint text, vision, audio, and action generation and reasoning in Physical AI.

Reinforcement learning system for in-hand cube reorientation on the WUJI Hand with full sim-to-real deployment

wuji-technology/wuji-mjlab1141d

Trains PPO policies for full-SO(3) in-hand cube reorientation on WUJI Hand1.0 in mjlab with GPU-batched MuJoCo physics and supplies a sim-to-real deployment bridge.

62

xAI co-founder Guodong Zhang clarifies that Grok Imagine was driven by imhaotian, disputing claims that Ethan He led the projectOther contributors endorsed prioritizing technical execution over post-hoc narratives

7h|33.4k25039

LQ#737|@LIANHUIQ

YS#603|@YING11231

71

Andreessen Horowitz launches global expansion initiatives, appointing former US official Anne Neuberger as Head of Global AffairsThe effort targets defense, AI, and cybersecurity tech partnerships.

28m|48.2k232134

SW#1814|@SARAHDINGWANG

ET#1253|@ERIKTORENBERG

81

NVIDIA reduces SOCAMM capacity for its upcoming Rubin architecture to address memory shortagesSemiAnalysis first reported the hardware design change.

2h|58.9k17349

PR#1332|@8TEAPI

DP#170|@DYLAN522P

90

ICRA 2026 awards Best Robot Learning Paper to research showing camera conditioning enables view-invariant robotic policy learningThe work reinforces the necessity of 3D geometric understanding.

3h|7.6k6830

KD#908|@CSPROFKGD

CP#732|@CHRIS_J_PAXTON

100

Gautam Kamath and co-authors win the 2026 Gödel Prize for their 2016 paper on high-dimensional robust estimatorsThe award is theoretical computer science's highest honor

2h|4.2k607

CS#409|@CSABASZEPESVARI

GK#210|@THEGAUTAMKAMATH

HL#35|@HUGO_LAROCHELLE

110

Nvidia-backed robotics startup Generalist AI raises $400 million at a $2 billion valuation to develop physical AI systemsThe startup released its GEN-1 physical-task model in April.

2h|15.1k4315

FR#1200|@FRASER

AC#91|@ACHOWDHERY

12

NVIDIA Pre-Trains Nemotron 3 Ultra 550B With NVFP4

59m|10.9k13918

WH#1430|@NREHIEW_

LA#980|@SCALING01

131

Helen Toner appointed permanent Executive Director of Georgetown's Center for Security and Emerging TechnologyToner, a CSET co-founder, previously served as interim director.

1h|3.1k643

CP#1092|@CHRISPAINTERYUP

141

Modal's Charles Frye shares his "sniff eval" prompt requiring LLMs to explain Singular Value Decomposition as a Harvard-educated pirateNVIDIA's Nemotron 3 Ultra solved it with a SEAS pun.

2h|2.1k253

B(#982|@ANDERSONBCDEFG

C🎉#863|@CHARLES_IRL

151

Nicolas Papernot's team demonstrates a conceptual AI worm that uses PrintNightmare and Docker exploits to compromise 25 of 33 hostsThe worm hijacks GPU compute to run open-weight LLMs.

7h|1.8k84

DJ#1988|@DAN_JEFFRIES1

NP#528|@NICOLASPAPERNOT

161

Gary Marcus argues the eventual winners of the AI transition will rely on neurosymbolic systemsThis responds to Carlos E. Perez's AI dot-com boom analogy.

3h|1.3k61

CE#1598|@INTUITMACHINE

GM#153|@GARYMARCUS

171

Gary Marcus argues semantic coherence is required for trustworthy AI, while Carlos Perez counters with a semiotic frameworkPerez argues language imposes order on intuitive AI models.

1h|20120

CE#1598|@INTUITMACHINE

GM#153|@GARYMARCUS

Github Stars

(7 days)

NVIDIA/cosmos1D AGO
Hosts omnimodal Mixture-of-Transformers models for joint text, vision, audio, and action generation and reasoning in Physical AI.
489
1452
260
8.7k stars
…technology/wuji-mjlab1D AGO
Trains PPO policies for full-SO(3) in-hand cube reorientation on WUJI Hand1.0 in mjlab with GPU-batched MuJoCo physics and supplies a sim-to-real deployment bridge.
1523
772
114 stars
…le-gemma/gemma-skills3D AGO
Supplies skills for Gemma model and agent interactions including a dev skill for app building.
927
486
267 stars
Yifei-Zuo/Parallax4D AGO
Implements parameterized local linear attention mechanisms for efficient language modeling.
260
235
46 stars

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

Yesterday's Top Stories, Jun 3, 2026.

Posts: 5,151|Clusters: 1,188|Stories frozen at midnight PT

10

Google releases Gemma 4 12B, an encoder-free multimodal model under Apache 2.0 that beats the larger Gemma 3 27BThe model runs locally on laptops with 16GB VRAM.

14h|4M33.1k#1BOOKMARKED11.5k

TM#1913|@ISCIENCELUVR

🎭#1807|@DEEPFATES

WH#1486|@NREHIEW_

DH#995|@DANIELHANCHEN

RP#852|@ROHANPAUL_AI

OS#800|@OSANSEVIERO

KA#635|@YACINEMTB

TA#472|@VICTORTAELIN

LB#451|@GIFFMANA

AJ#449|@ARMANDJOULIN

👩‍💻P#428|@DYNAMICWEBPAIGE

PT#250|@ELDER_PLINIUS

SP#10|@SUNDARPICHAI

DH#4|@DEMISHASSABIS

20

Ideogram founder Mohammad Norouzi releases Ideogram 4.0, an open-weight text-to-image model supporting local inference and custom fine-tuningThe model is available immediately across all subscription tiers.

1d|1.5M7.6k5.7k

AC#1823|@ANDREW_N_CARR

SR#1376|@CLONEOFSIMO

LI#698|@LISHALI88

JM#578|@VENTURETWINS

MA#169|@MARTIN_CASADO

RS#156|@SCOBLEIZER

AK#123|@_AKHALIQ

EM#29|@EMOSTAQUE

30

xAI releases Grok Imagine 1.5 Preview on its API and starts rolling it out to standard chatbot subscribersElon Musk showcased the update with an AI-generated Trojan War trailer.

13h|#1VIEWS44.5M#1LIKED81.8k9.5k

AC#406|@ANDREWCURRAN_

B(#243|@BEFFJEZOS

SJ#220|@FUTUREJURVETSON

@L#82|@LEVELSIO

40

Town exits beta and raises $55 million Series A led by a16z for its administrative AI assistantThe assistant requires explicit user approval before executing workflows

16h|869.2k1.9k1.1k

SW#1134|@SARAHDINGWANG

JM#578|@VENTURETWINS

MB#541|@MATTHEWBERMAN

ET#236|@ERIKTORENBERG

RS#156|@SCOBLEIZER

SG#92|@SARANORMOUS

BE#48|@BHOROWITZ

50

OpenAI releases a promotional video teasing its Codex model but provides no technical details or release plansThe "It's time to fly" clip sparked online speculation

11h|1.7M11.2k1.9k

JA#1487|@JXNLCO

SW#1148|@SHERWINWU

60

Lassie raises $47 million led by a16z to automate administrative paperwork for dental and medical practicesThe software currently saves 700 practices 30 hours monthly.

16h|1.5M4.8k4.4k

OM#688|@OMOORETWEETS

AK#279|@ALEXRKONRAD

RS#156|@SCOBLEIZER

@L#82|@LEVELSIO

70

OpenAI proposes a federal frontier AI safety framework that mandates model evaluations but preempts state-level safety lawsThe proposed regulator would lack power to block model deployments.

13h|310.8k1.9k375

BB#641|@BOAZBARAKTCS

AC#406|@ANDREWCURRAN_

80

Miso releases Miso One, an 8B open-weights text-to-speech model with 110ms latency and one-shot voice cloningThe open weights on GitHub allow for local self-hosting.

14h|1.4M5.9k6.8k

JL#402|@BYJOHNNYLEE

RS#156|@SCOBLEIZER

90

Ted Chiang argues text outputs drive AI consciousness claims, while structurally similar models like AlphaFold and Sora are ignoredCommenters blame marketing hype for the public's anthropomorphization of software.

9h|232.1k3.2k585

KA#635|@YACINEMTB

J⧉#225|@REPLIGATE

NS#185|@NOAHPINION

100

Reve launches Reve 2.0, a 4K image generator using layout-based code control that ranks second on Text-to-Image ArenaThe model ranks just behind OpenAI's gpt-image-2

11h|7.9M3.2k2.6k

KK#1477|@KEVINAKWOK

RS#156|@SCOBLEIZER

110

Harvey and Fireworks AI hybrid agent routing outperforms Claude Opus 4.7 on legal benchmarks at 61% lower costThe hybrid setup beat GPT-5.5, which scored 11% at $560.

13h|399.6k1.4k1.1k

C🎉#1058|@CHARLES_IRL

TW#89|@THOM_WOLF

C🤗#41|@CLEMENTDELANGUE

120

Sci-fi author Ted Chiang argues that artificial intelligence systems are not consciousThe essay prompted debate over Chiang's recent AI commentary.

14h|84.7k827234

CP#1889|@CHRISGPOTTS

CP#820|@CHRIS_J_PAXTON

MB#70|@MILES_BRUNDAGE

130

Pro-AI super PAC backed by Greg Brockman and Marc Andreessen ran fake anti-AI sockpuppet accounts on X, investigation findsThe sockpuppets promoted an affiliate called Build American AI.

15h|167.6k1.2k161

DF#669|@DANIELLEFONG

140

Sasha Rush explains targeted on-policy self-distillation, a reinforcement learning technique that corrects specific LLM rollout errorsThe method bypasses noisy end-of-trajectory final rewards.

4h|143.8k1.8k1.6k

WH#1486|@NREHIEW_

SR#940|@SRUSH_NLP

NL#876|@NATOLAMBERT

DP#106|@DWARKESH_SP

150

Cloudflare's Matthew Prince claims agentic bot traffic has surpassed human traffic, sparking debate over HTML-specific metricsOverall HTTP requests remain 69.6% human, according to Radar data.

9h|169.4k1.1k125

AL#695|@ALFRED_LIN

MP#455|@EASTDAKOTA

👩‍💻P#428|@DYNAMICWEBPAIGE

BS#274|@BILAWALSIDHU

EB#64|@ERIKBRYN

160

Ashok Elluswamy, Tesla's VP of AI Software, announces the launch of unsupervised Robotaxi service in Austin, TexasVehicles operate without human oversight inside the geofenced territory.

14h|2.1M14.5k396

AE#573|@AELLUSWAMY

RS#156|@SCOBLEIZER

170

Microsoft AI announces MAI-Thinking-1 and MAI-Code-1-Flash, claiming they outperform Claude on coding and reasoning benchmarksMAI-Code-1-Flash will integrate exclusively inside GitHub Copilot.

1d|28.1k24646

SA#1035|@PRFSANJEEVARORA

SR#940|@SRUSH_NLP

RP#852|@ROHANPAUL_AI

DR#422|@DELIPRAO

ND#130|@NANDODF

180

Gopuff launches Go, an AI shopping assistant powered by xAI's Grok Voice, Imagine, and Reasoning APIsxAI's Eric Jiang says Go handles automated ordering.

15h|11.1M21.1k2k

RS#156|@SCOBLEIZER

190

OpenAI resets Codex API usage limits across paid plans after three outages in 24 hoursExcessive request volumes triggered the platform's rate-limiting system

1d|548.4k9.1k389

JA#1487|@JXNLCO

MB#541|@MATTHEWBERMAN

TI#117|@THSOTTIAUX

200

Keith Rabois and Aaron Levie cite historic US job openings surge to argue AI is boosting software engineering demandApril job openings surged by 731,000, led by professional services.

18h|381.8k1.9k363

DJ#979|@DAN_JEFFRIES1

GM#177|@GARYMARCUS

210

Applied Compute's Yash Patil argues enterprise transition to self-hosted open-weight models is inevitableDeepSeek is one of Ramp's fastest-growing AI vendors.

16h|163.5k643240

ZE#1476|@ZEPHYR_Z9

FC#803|@ALTIMOR

T(#789|@TEORTAXESTEX

YP#650|@YPATIL125

220

Sam Altman, Dario Amodei, and Demis Hassabis urge Congress to mandate screening of synthetic nucleic acids to prevent AI biosecurity risksThe proposed rules also target physical DNA manufacturing equipment.

5h|248.4k2.6k862

LA#1524|@SCALING01

AC#406|@ANDREWCURRAN_

230

CVPR 2026 workshop "Bitter Lessons in Computer Vision" to debate scaling laws versus hand-crafted engineeringJon Barron will analyze scaling principles in 3D vision.

1d|76.4k925348

AB#1507|@ABURSUC

GG#1124|@GEORGIAGKIOXARI

240

MechanizeWork seeks to hire a puzzle maker to design evaluation challenges that current LLMs cannot solveNeedle in a Haystack creator Greg Kamradt shared the listing.

13h|148.4k2.9k212

AC#1823|@ANDREW_N_CARR

🎭#1807|@DEEPFATES

GK#459|@GREGKAMRADT

MB#70|@MILES_BRUNDAGE

250

Candy急兒 launches Monako Glass, a wearable Linux computer built into eyeglasses designed to run AI coding agentsThe eyeglasses support mobile vibe-coding in diverse environments.

16h|612.9k4.8k2.9k

RS#156|@SCOBLEIZER

260

Ideogram launches Ideogram 4.0, its first open-weight text-to-image model supporting local inference and fine-tuningThe model is also accessible via the company's API.

14h|172.4k1.3k642

FO#1390|@FOFRAI

MA#169|@MARTIN_CASADO

270

OpenAI's roon jokes that his Waymo self-driving car is conscious, prompting playful debate over machine sentienceCommenters debated if corporate systems are inherently misaligned.

7h|104.1k1.7k70

JL#1959|@ARTIRKEL

AC#406|@ANDREWCURRAN_

280

Cloudflare partners with xAI to integrate Grok LLMs and multimodal video models directly into its AI GatewayUnified billing eliminates the need for separate xAI API keys

9h|5.7M14.7k754

290

Factory launches Factory Router, an automated model routing system designed to cut AI inference costs by 25%The system dynamically routes tasks to smaller, cheaper models.

1d|118.8k59740

RP#852|@ROHANPAUL_AI

SM#701|@SHAUNMMAGUIRE

PG#570|@GRADYPB

300

Bain survey finds corporate AI investments yield disappointing financial returns and fail to deliver measurable business valueSimilar reports from McKinsey and MIT also found weak returns

1d|117.4k58895

CP#820|@CHRIS_J_PAXTON

GM#177|@GARYMARCUS

EM#96|@EMOLLICK

GT#23|@GARRYTAN