AI News

Today's Highlights

#1 last 24hrs

SpaceX S-1 filing reveals Google will pay $920 million monthly to lease 110,000 Nvidia GPUs

B(#850|@BEFFJEZOS

CP#725|@CHAMATH

AC#518|@ANDREWCURRAN_

NB#242|@NATHANBENAICH

EM#169|@EMOSTAQUE

#1 BOOKMARKED

3.9k

Markus J. Buehler proposes a category-theory framework that lets AI scientists dynamically expand their reasoning schema

RS#323|@SCOBLEIZER

DR#56|@ROYDANROY

FASTEST CLIMBING

ML researcher @kalomaze says a KV cache bug caused transformer inference routing errors that exposed model outputs to other users

CH#1448|@KIMMONISMUS

LB#980|@REMEMBERLENNY

KA#839|@KALOMAZE

RISING LIKES

277

DeepDream creator Alex Mordvintsev teases a rotating 3D wireframe visualization with no accompanying technical details

Top Stories

Posts:—|Clusters:—|Next Crawl:—

10

SpaceX S-1 filing reveals Google will pay $920 million monthly to lease 110,000 Nvidia GPUsSpaceX's AI cloud run rate now reaches $26 billion.

7h|#1VIEWS4.7M#1LIKED25.9k2.8k#1COMMENTS2k

XL#1671|@XLR8HARDER

AA#1559|@AUSTEN

ZE#1471|@ZEPHYR_Z9

BI#1366|@BILALTWOVEC

SM#1342|@SHAUNMMAGUIRE

RO#1214|@KRISHNANROHIT

SI#1096|@SIGNULLL

CP#1062|@CHRISPAINTERYUP

RP#1031|@ROHANPAUL_AI

B(#850|@BEFFJEZOS

CP#725|@CHAMATH

AC#518|@ANDREWCURRAN_

NB#242|@NATHANBENAICH

EM#169|@EMOSTAQUE

20

Google DeepMind releases Gemma 4 QAT, but Unsloth developer Daniel Han warns naive llama.cpp conversions suffer accuracy lossThe release shrinks Gemma 4 E2B's footprint to 1GB.

9h|437.9k5k2.2k

CH#1448|@KIMMONISMUS

UT#1430|@UTKUEVCI

DH#773|@DANIELHANCHEN

YS#608|@YING11231

OS#486|@OSANSEVIERO

👩‍💻P#270|@DYNAMICWEBPAIGE

30

XPRIZE founder Peter Diamandis argues universities charge up to $300,000 for degrees teaching skills now available via free LLMsHe calls this discrepancy a significant misallocation of capital.

11h|1M11k827

B(#850|@BEFFJEZOS

HP#459|@HYHIEU226

CL#438|@CHENGLEISI

EM#74|@ELONMUSK

40

Senator Bernie Sanders proposes the US government acquire a 50 percent equity stake in major AI companiesGary Marcus warns state ownership risks establishing social credit.

7h|707.5k3.6k1.1k

B(#850|@BEFFJEZOS

DW#390|@DEANWBALL

GM#157|@GARYMARCUS

51

Markus J. Buehler proposes a category-theory framework that lets AI scientists dynamically expand their reasoning schemaThe system uses typed copresheaves to mathematically quantify novelty.

12h|577.5k3.1k#1BOOKMARKED3.9k

RS#323|@SCOBLEIZER

DR#56|@ROYDANROY

New github stars (48hrs)

Benchmark measuring continual learning of AI agents over multi-episode shared environments

pgasawa/continual-learning-bench1271d

Benchmarks AI agents on learning from repeated interactions across multi-episode tasks with adaptation metrics.

Sandboxes, SDKs and benchmarks for computer-use agents that control full macOS, Windows and Linux desktops

trycua/cua17.7k1d

Supplies sandboxes, SDKs, and benchmarks for AI agents to control full desktops on macOS, Linux, and Windows.

Interactive terminal agent using DSPy recursive language models for iterative Python coding

diego-lima/rlmy91d

61

Google DeepMind, UCL, and Oxford team wins CVPR 2026 Best Paper for dynamic scene reconstruction method D4RTThe paper was selected from 74 shortlisted candidates.

10h|57.5k731172

DD#1728|@DIMADAMEN

ML#1665|@MARIOLUCIC_

AB#1577|@ABURSUC

KD#932|@CSPROFKGD

71

President Trump signs directive ordering U.S. military and intelligence agencies to deploy commercial AI on classified networksThe order updates autonomous weapons guidance and mandates compute facilities

6h|121k1.1k136

LH#1370|@OHLENNART

BA#1291|@BAYESLORD

DK#849|@DKAUSHIK96

81

Anysphere co-founder Michael Truell launches Design Mode for Cursor, enabling UI editing through voice, drawing, and live gesturingThe feature aims to replace text-only AI agent prompts.

7h|595k3k312

MT#506|@MNTRUELL

EM#74|@ELONMUSK

93

Sakana AI launches its Recursive Self-Improvement Lab to build autonomous, self-improving AI systemsIt unifies existing company projects like The AI Scientist.

8h|134.1k1.2k442

RL#1463|@ROBERTTLANGE

B(#850|@BEFFJEZOS

HA#18|@HARDMARU

100

Y Combinator co-founder Paul Graham argues big corporations' failure to profit from LLM tokens is a predictable adoption phaseGarry Tan says organizational skill gaps hamper corporate execution.

12h|258k2.7k406

GT#266|@GARRYTAN

116

Anthropic, OpenAI, and other labs disclose active initiatives for recursive self-improvementAnthropic says its Claude model is already accelerating development.

9h|33.3k14282

JW#1469|@JIAXINWEN22

KR#1055|@KYLIEBYTES

AA#669|@ARYAMAN2020

CL#438|@CHENGLEISI

121

OpenAI's Vaibhav Srivastav apologizes after a billing system glitch accidentally suspended customer accountsAffected users feared bans for discussing competitor products.

9h|347.8k2.7k243

V(#1853|@REACH_VB

130

Open gaming foundation model NITROGEN and Meta's SAM 3D earn Best Paper Honorable Mentions at CVPR 2026NITROGEN was trained on 40,000 hours of gameplay.

9h|27.4k26459

YW#1595|@YUEWANG314

GS#1523|@GUANYASHI

GG#679|@GEORGIAGKIOXARI

JF#31|@DRJIMFAN

140

Anthropic says Claude Opus 4.7 matches dedicated chemistry software like MestReNova on predicting molecular structures from NMR spectraIt also reconstructs molecular structures directly from NMR spectra

5h|123.8k1.9k656

ZE#1471|@ZEPHYR_Z9

RP#1031|@ROHANPAUL_AI

150

Atari Lab demonstrates advanced humanoid robot locomotion and multi-step planning under its MotionDisco projectDemos show the robot climbing desks and balancing on boxes.

14h|119.1k703208

CP#737|@CHRIS_J_PAXTON

T(#421|@TEORTAXESTEX

160

Independent AI researcher Pliny the Liberator releases Enthea, a single-file HTML psychedelic visualizer built with ClaudeThe tool uses mathematical models of the visual cortex.

10h|30.1k629366

PT#640|@ELDER_PLINIUS

RS#323|@SCOBLEIZER

176

ML researcher @kalomaze says a KV cache bug caused transformer inference routing errors that exposed model outputs to other usersThe failure mode resembles OpenAI's 2023 routing errors

7h|76.2k71398

FASTEST CLIMBING+6

CH#1448|@KIMMONISMUS

LB#980|@REMEMBERLENNY

KA#839|@KALOMAZE

181

Nikolay Savinov joins OpenAI's London office as its pretraining unit seeks new researchers to expandSavinov will focus on pretraining and long-context research.

8h|53.3k387102

SL#1022|@SAMUELMLSMITH

AC#240|@_AIDAN_CLARK_

SB#50|@SEBASTIENBUBECK

191

Sara Hooker's AI startup Adaption launches a $50,000 challenge to build models using its AutoScientist workflow acceleratorThe four-week competition features ten categories, including medicine.

12h|33.6k216219

NI#305|@NILOOFAR_MIRE

SH#14|@SARAHOOKR

200

Perplexity integrates NVIDIA's open-source Nemotron 3 Ultra model to power long-running agentic workflows on its Pro and Max tiersThe model is live on the platform's Computer feature.

6h|77k728141

SD#1794|@INDUCTIONHEADS

AS#120|@ARAVSRINIVAS

Rising Stories

KV Cache Bug7H AGO
ML researcher @kalomaze says a KV cache bug caused transformer inference routing errors that exposed model outputs to other users
#839
27420.6k
Vortex Inference Win5H AGO
InfiniAI Lab releases Vortex, an agent-designed sparse attention framework that accelerates LLM inference throughput by up to 4.7x
#534
132.5k
Reve 2.0 Debuts7H AGO
Reve 2.0 debuts at #2 in LMSYS Image Arena as three new models enter the top 10
#471
1435.4k
Dreaming Memory Rollout7H AGO
OpenAI deploys dreaming memory system for ChatGPT to actively update user profiles across conversations
#1031
15612.7k
Seinfeld Job Zine6H AGO
Tim Hwang uses Claude to map every occupation depicted across all 174 Seinfeld episodes for a new labor zine
#20
855.8k

Recent Stars

maria-rcks/mode38M AGO
Applies minimal patches via bun scripts to customize T3 Code into a preferred dev environment.
1784
5 stars
ogulcancelik/herdr1H AGO
Multiplexes AI agents in the terminal with workspaces, tabs, panes, state tracking, and tmux-like persistence.
1784
4.6k stars
mikayla-maki/loom2H AGO
Manages and runs AI agent harnesses from TOML manifests specifying harnesses, layered sessions, tools, and capability grants for security.
1784
9 stars
wangmiaowei/BiMotion2H AGO
Generates text-conditioned dynamic 3D character animations by representing motions as differentiable B-spline curves.
260
43 stars
…bookresearch/meshflow2H AGO
Generates artistic meshes from input geometry and optional reference images via MeshVAE encoding and flow-matching DiT.
260
6 stars

Github Stars

(7 days)

…tinual-learning-bench1D AGO
Benchmarks AI agents on learning from repeated interactions across multi-episode tasks with adaptation metrics.
862
846
127 stars
trycua/cua1D 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/rlmy1D AGO
UNKNOWN.
862
158
9 stars

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