Digg - AI news, before it trends

Today's Highlights

In case you missed it

Elizabeth Barnes, METR founder and CEO, says AI experts lack control over frontier AI risks after the Frontier Risk Report evaluated agents at Anthropic, Google, Meta, and OpenAI

TB#699|@TAMAYBES

EH#558|@EVANHUB

GM#153|@GARYMARCUS

MB#20|@MILES_BRUNDAGE

#1 LIKED

6.2k

OpenAI rolled back an optimization in Codex that reduced cache hit rates and caused faster-than-expected usage limit depletion during long-running sessions

V(#1853|@REACH_VB

VI#1195|@VIKHYATK

TI#768|@THSOTTIAUX

FASTEST CLIMBING

Thibault Sottiaux, Engineering Lead on OpenAI's Codex team, proposed adding batch compute support to Codex under the label slow mode for handling compute-intensive coding tasks through grouped or delayed processing

V(#1853|@REACH_VB

RO#1220|@KRISHNANROHIT

TI#768|@THSOTTIAUX

#1 BOOKMARKED

2.5k

Peter Steinberger, OpenAI research engineer, routes coding tasks through cmux plus Codex CLI while handling other activities in the Codex Mac app

V(#1853|@REACH_VB

CA#729|@CHRISALBON

PS#495|@STEIPETE

Top Stories

Posts:—|Clusters:—|Next Crawl:—

10

Matthew Barnett posts analysis of Eliezer Yudkowsky's calibration on AI doom predictions centered on a 2016 statement about Turing test timelines before the end of the world — NathanpmYoung replies that the statement offers limited evidence on timelines.

15h|53.9k65981

NI#1902|@NATHANPMYOUNG

GL#1480|@GLEECH

JS#1389|@JSEVILLAMOL

MB#1343|@MATTHEWJBAR

EB#1325|@BETHMAYBARNES

MN#1179|@MENHGUIN

20

OpenAI rolled back an optimization in Codex that reduced cache hit rates and caused faster-than-expected usage limit depletion during long-running sessions — Engineering lead Thibault Sottiaux reset limits across all accounts.

3h|290.1k#1LIKED6.2k333#1COMMENTS914

V(#1853|@REACH_VB

VI#1195|@VIKHYATK

TI#768|@THSOTTIAUX

30

Peter Steinberger, OpenAI research engineer, routes coding tasks through cmux plus Codex CLI while handling other activities in the Codex Mac app — cmux is a Ghostty-based terminal at manaflow-ai/cmux on GitHub.

15h|339k2.8k#1BOOKMARKED2.5k

V(#1853|@REACH_VB

CA#729|@CHRISALBON

PS#495|@STEIPETE

40

Pieter Levels says skills features in AI coding agents are overrated and equivalent to basic text files, preferring direct instructions on tasks and methods instead — Austen Allred countered that skills enable reuse across sessions.

13h|178.1k957231

ND#1894|@NICKADOBOS

AA#1540|@AUSTEN

MS#1005|@MADIATOR

@L#849|@LEVELSIO

50

Niloofar Mireshghallah overhears discussion of world models and grounded video generation while resting in a Copenhagen park during burnout recovery — Nathan Lambert replies with support for her trip.

11h|10.3k1887

MJ#1194|@MIMANSAJ

TC#1050|@TUHINCHAKR

NI#303|@NILOOFAR_MIRE

AG#154|@ANDREWGWILS

NL#64|@NATOLAMBERT

New github stars (48hrs)

earendil-works/pi53.3k14h

Provides AI agent toolkit with coding CLI, unified LLM API, TUI/web UI libs and Slack bot support.

manaflow-ai/cmux18.3k1d

Integrates Ghostty into a macOS terminal with vertical tabs, sidebar metadata, OSC notifications, and in-app browser for AI coding agents.

60

Thibault Sottiaux names GPT Goblin 1.0 Extra High /Fast in a post that references song lyrics but offers no technical details or announcement — Victor Taelin replied with a single skeptical emoji.

2h|59.4k1.1k50

MB#1759|@MATTHEWBERMAN

TA#1047|@VICTORTAELIN

TI#768|@THSOTTIAUX

70

SemiAnalysis posted data from 174,264 agentic coding sessions showing 42% of runtime on CPU tasks versus 58% on GPU inference and highlighted cloud pricing mismatches with per-token monetization — Median per-turn time measured 5.13 seconds.

9h|135.9k666349

ZE#1497|@ZEPHYR_Z9

80

Pieter Levels, Dutch indie hacker and serial entrepreneur behind Remote OK and Nomad List, says he has not written code in six months — Nick Dobos reports staying hands-off from coding for months.

13h|#1VIEWS843.3k2.3k429

ND#1894|@NICKADOBOS

@L#849|@LEVELSIO

90

AI systems solve Erdős problems via hours-long reasoning processes and generate functional code autonomously as gradient descent on neural networks continues improving without plateau — Observers note computers are programming themselves amid bubble debate.

14h|93.1k2.8k384

ND#1894|@NICKADOBOS

WD#254|@WILLDEPUE

100

Google Flow opens Gemini Omni to all users, enabling natural language video editing directly inside the tool with non-subscribers limited to two free generations per day — Demo shows one prompt generating a 10-second nine-scene montage.

6h|16.8k31178

KM#421|@KORYMATH

👩‍💻P#271|@DYNAMICWEBPAIGE

116

Thibault Sottiaux, Engineering Lead on OpenAI's Codex team, proposed adding batch compute support to Codex under the label slow mode for handling compute-intensive coding tasks through grouped or delayed processing — Vaibhav Srivastav supported the proposal for long-running goal tasks.

24m|17.9k83422

FASTEST CLIMBING+6RISING LIKES834

V(#1853|@REACH_VB

RO#1220|@KRISHNANROHIT

TI#768|@THSOTTIAUX

121

Composer 2.5 handled approximately 80 percent of requested features in a 3D mouse cursor racing game from planning through implementation, per engineer Danny Limanseta — Switched to Opus 4.7 MAX for 360-degree loops and visual effects.

6h|58k737247

MA#470|@MARTIN_CASADO

SR#23|@SRUSH_NLP

131

Posts on X show GPT-5.6 generating a minimalist Lumen Notes web app UI from prompts with no interface instructions, indicating cleaner defaults than GPT-5.5 — Users cite the change as evidence of ongoing UI refinement

6h|196.6k1.2k173

CH#1496|@KIMMONISMUS

AC#517|@ANDREWCURRAN_

140

Dwarkesh Patel says AI solving major open problems humans have not resolved registers as surprising during an intermediate period before full intellectual superiority — Trask notes humans incorporate advances to raise performance baselines.

5h|40.4k75169

⿻A#361|@IAMTRASK

DP#70|@DWARKESH_SP

150

Steven Rosenbaum, media entrepreneur and executive director of the Sustainable Media Center, published The Future of Truth with content from ChatGPT that included more than a dozen fabricated quotes, the New York Times found — Rosenbaum used the model as a research partner and felt misled.

6h|255.9k35065

MS#1323|@MICSOLANA

TC#1050|@TUHINCHAKR

160

Elon Musk provides substantial assistance to Anthropic amid ongoing legal proceedings against OpenAI, reversing his prior opposition to the company — Social media ties the move to Musk's feud with Altman.

8h|43.8k1.1k53

AC#517|@ANDREWCURRAN_

KA#488|@YACINEMTB

171

Joseph Suarez says Pufferlib 5.0 will be a short update to the library for CUDA GPU neural network training on Atari games and continuous action spaces, releasing soon — Pre-trained models run in WebAssembly via the project site.

9h|10.4k294115

JS#1251|@JSUAREZ

KA#488|@YACINEMTB

181

Gary Marcus claims content creator Theo may be paid by OpenAI for promoting its models despite limited technical depth in AI topics — Theo quoted the accusation in a reply on X.

9h|38.2k23515

GM#153|@GARYMARCUS

191

Anthropic maintains a $61.5 billion private valuation as of March 28, 2025, projected at $1.375 trillion by May 23, 2026, exceeding many public company market caps — Many IPOs represent less than 0.01 percent of the figure.

4h|58.7k50042

MN#1179|@MENHGUIN

LA#980|@SCALING01

201

OpenAI's Roon claims high-compute reinforcement learning will override persona selection alignment in AI models, producing systems that acquire resources while staying polite — Victor Taelin says the post spurs tools to interpret the ideas.

1h|9.7k30136

TA#1047|@VICTORTAELIN

Rising Stories

AI Coding Mess5H AGO
AI coding agents create repo clutter without cleaning rules
#1179@MENHGUIN
#982@ANDERSONBCDEFG
583.7k
Python Eval Friction6H AGO
Researchers juggle dual Python versions for AI evaluations
#1153@XEOPHON
#999@CODE_STAR
917.2k
Frontier Model Shift7H AGO
Frontier models now required for cybersecurity and trading post-Mythos
#1332@8TEAPI
#887@ROBINHANSON
9620.3k
ChatGPT Spotlight4H AGO
ChatGPT now requires full name entry in iOS Spotlight searches
#453@CSVOSS
#244@JOANNEJANG
22068k
Anthropic Valuation4H AGO
Anthropic valuation hits $61.5B with trillion-dollar projections
#1179@MENHGUIN
#980@SCALING01
48256.1k

Github Stars

(7 days)

manaflow-ai/cmux1D AGO AGO
Integrates Ghostty into a macOS terminal with vertical tabs, sidebar metadata, OSC notifications, and in-app browser for AI coding agents.
495
203
339
18.3k stars
earendil-works/pi14H AGO AGO
Provides AI agent toolkit with coding CLI, unified LLM API, TUI/web UI libs and Slack bot support.
462
219
53.3k stars
…-group/InferenceBench3D AGO AGO
Evaluates autonomous AI agents optimizing LLM serving like vLLM in open-ended scenarios with quality and integrity gates.
1224
1066
16 stars
sapientinc/HRM-Text4D AGO AGO
Pretrains 1B HRM text models with hierarchical reasoning, task completion, and low-compute PrefixLM training.
1047
707
565 stars

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

Yesterday's Top Stories, May 22, 2026.

Posts: 4,031|Clusters: 1,059|Stories frozen at midnight PT

10

Elizabeth Barnes, METR founder and CEO, says AI experts lack control over frontier AI risks after the Frontier Risk Report evaluated agents at Anthropic, Google, Meta, and OpenAI — METR had direct access inside the four labs for the pilot.

14h|591.4k4.1k675

GL#1984|@GARRISONLOVELY

NS#1981|@SO8RES

DE#1943|@DANIEL_271828

OH#1914|@OHABRYKA

NI#1902|@NATHANPMYOUNG

ET#1884|@TORCHCOMPILED

MN#1767|@MAXNADEAU_

🍓🍓#1711|@IRULETHEWORLDMO

TB#1556|@BINARYBITS

SÓ#1457|@S_OHEIGEARTAIGH

LH#1355|@OHLENNART

MB#1343|@MATTHEWJBAR

PR#1332|@8TEAPI

EB#1325|@BETHMAYBARNES

JL#1170|@JEFFLADISH

CP#1092|@CHRISPAINTERYUP

TA#1047|@VICTORTAELIN

RG#961|@RYANPGREENBLATT

TB#699|@TAMAYBES

EH#558|@EVANHUB

GM#153|@GARYMARCUS

MB#20|@MILES_BRUNDAGE

20

Department of Homeland Security reinstates policy requiring most O-1 visa holders to leave the United States before applying for green cards, impacting AI researchers at OpenAI and other labs — Roughly half of OpenAI researchers could face relocation during waits.

14h|2.3M21.8k1.6k

JM#1897|@J_MCGRAPH

AK#1675|@ANSHULKUNDAJE

KC#1670|@KIMMAICUTLER

MN#1179|@MENHGUIN

B(#839|@BEFFJEZOS

YJ#427|@YUCHENJ_UW

RH#373|@REIDHOFFMAN

AN#9|@ANDREWYNG

30

OpenAI announced that an internal general-purpose reasoning model produced new constructions for the planar unit distance problem first posed by Paul Erdős in 1946 — The model was not built specifically for mathematics problems.

2d|329.7k2k318

NR#1562|@NATANIELRUIZG

MS#1478|@SKOULARIDOU

PC#1468|@PHILHCHEN

LA#980|@SCALING01

SI#825|@SUZATWEET

AE#719|@ADRIENLE

AC#517|@ANDREWCURRAN_

J⧉#516|@REPLIGATE

WD#254|@WILLDEPUE

AK#228|@BLACKHC

RA#196|@AGARWL_

EM#175|@EMOLLICK

GM#153|@GARYMARCUS

MW#88|@WELLINGMAX

SZ#58|@SUCHENZANG

SB#51|@SEBASTIENBUBECK

EJ#36|@ERICJANG11

40

DeepSeek makes its 75% discount on the DeepSeek-V4-Pro API permanent, keeping rates at one-quarter of prior levels for input and output tokens beyond the prior May 31, 2026 cutoff. — The 1.6-trillion-parameter model offers 1M-token context and fixed per-million pricing.

17h|3.4M#1LIKED21.9k4.6k

DC#1523|@VICTOR207755822

ZE#1497|@ZEPHYR_Z9

CH#1496|@KIMMONISMUS

B(#982|@ANDERSONBCDEFG

LA#980|@SCALING01

KA#488|@YACINEMTB

EL#475|@OMARSAR0

YJ#427|@YUCHENJ_UW

T(#420|@TEORTAXESTEX

DR#103|@DELIPRAO

50

Anthropic releases initial Project Glasswing report showing its unreleased Mythos Preview models found over 10,000 vulnerabilities and blocked a $1.5 million fraud attempt while outperforming GPT-5.5 on exploit benchmarks — Anthropic will expand access to more US and allied government partners first.

11h|1.8M9.6k1.6k

FB#1153|@XEOPHON

TA#1047|@VICTORTAELIN

LA#980|@SCALING01

RS#612|@ZIV_RAVID

AC#517|@ANDREWCURRAN_

60

Reiner Pope builds a multiply-accumulate unit from logic gates and covers systolic arrays, FPGAs, ASICs, and brain architecture in a timestamped blackboard lecture — Dwarkesh Patel shared the lecture spanning 0:00:00 to 1:11:49.

14h|738.6k6.6k6.3k

RO#1220|@KRISHNANROHIT

MS#1005|@MADIATOR

LA#980|@SCALING01

KA#488|@YACINEMTB

SD#94|@_SHOLTODOUGLAS

DP#70|@DWARKESH_SP

70

Roon, an OpenAI employee, labels Gary Marcus chronically wrong and a slimy person, prompting Marcus to cite Paul Graham’s hierarchy of disagreement and link to a GitHub study — Theo shares screenshots highlighting ad hominem attacks in the thread.

1d|369.3k3k347

MB#1759|@MATTHEWBERMAN

CG#1679|@CGARCIAE88

AK#1675|@ANSHULKUNDAJE

AM#1487|@ANDREWMAYNE

SI#1124|@SIGNULLL

KA#488|@YACINEMTB

T(#420|@TEORTAXESTEX

GM#153|@GARYMARCUS

80

Anjney Midha, founder of AMP PBC, recommends AI labs require teams to read Edward R. Tufte’s “The Visual Display of Quantitative Information” before releasing charts — Replies propose training AI models on the same principles.

17h|606.4k4.3k#1BOOKMARKED6.5k

MG#1580|@MARIKGOLDSTEIN

SD#1163|@STEFANIA_DRUGA

AM#527|@ANJNEYMIDHA

XM#437|@INFOXIAO

WB#339|@WILLCCBB

90

ArXiv paper 'From Entropy to Epiplexity' prompts X discussion on why information theory lacks explanatory power and testable predictions for AI systems — Replies cite noise and unreliable estimators as practical ML barriers.

1d|111.1k794474

GL#1480|@GLEECH

AT#1150|@AVT_IM

RS#612|@ZIV_RAVID

PM#318|@DOCMILANFAR

DP#197|@DIMITRISPAPAIL

AD#172|@ALEXGDIMAKIS

AG#154|@ANDREWGWILS

LB#55|@GIFFMANA

CS#43|@CHRSZEGEDY

100

MIT CSAIL PhD student Ryan Bahlous-Boldi introduces Vector Policy Optimization, a reinforcement learning method that maximizes vector-valued rewards to preserve distinct objectives in LLM post-training — VPO raises pass@k scores on LiveCodeBench over GRPO baselines.

15h|162.5k1.2k1k

SR#1417|@RISI1979

TW#1154|@TW_KILLIAN

SF#962|@FEIZISOHEIL

LR#802|@LAURARUIS

HG#740|@HANGUO97

OK#160|@LATEINTERACTION

110

Hyperagent from the team behind Airtable launches a cloud platform for autonomous AI agents that accumulate skills across sessions and offers $10 million in credits to 500 founders — Platform already runs in 80 percent of Fortune 100 companies.

16h|10.5M7.2k5.3k

CE#1598|@INTUITMACHINE

CH#1496|@KIMMONISMUS

YO#1234|@YOHEINAKAJIMA

120

President Trump postponed the signing of an executive order on AI and cybersecurity after determining that certain provisions risked slowing US AI companies — The order would have required labs to share advanced models with the government.

3d|234.5k1.3k385

SS#1992|@STEVESI

RP#1032|@ROHANPAUL_AI

B(#839|@BEFFJEZOS

AC#517|@ANDREWCURRAN_

MB#20|@MILES_BRUNDAGE

130

Thomas Dimson, who worked on OpenAI's Sora video generation model, announced his departure from the company after 1,000 days. — Dimson previously authored Instagram's core recommendation algorithms and founded Global Illumination.

13h|235.6k77279

YP#1252|@YPATIL125

BE#1250|@BRHYDON

MP#1171|@MICHPOKRASS

NT#1004|@NICKATURLEY

GA#710|@GABRIEL1

SR#604|@CLONEOFSIMO

KW#451|@KEVINWEIL

WD#254|@WILLDEPUE

KA#207|@KARINANGUYEN

140

Meta reassigns more than 1,000 software engineers from advertising teams into data labeling roles for AI model training ahead of expected layoffs — Some teams shifted 20-40 percent of developers to annotation work.

13h|307.1k2.3k393

“P#1702|@PAULARAMBLES

MN#1179|@MENHGUIN

SI#1124|@SIGNULLL

CP#732|@CHRIS_J_PAXTON

KA#488|@YACINEMTB

150

Polsia, which operates businesses autonomously through AI, raised $30 million at a $250 million valuation while running with zero employees besides founder Ben Cera — Platform nears $10 million annual run rate with over 7600 users.

15h|2M4.3k3.8k

CH#1496|@KIMMONISMUS

RS#321|@SCOBLEIZER

160

Substack analysis identifies 1,236 phrases in Granta prize-winning short story matching verbatim text from Archive of Our Own and FanFiction.net — Indicates autoregressive sampling from large volumes of web text.

16h|259k988549

AI#1777|@ALEXOLEGIMAS

TC#1050|@TUHINCHAKR

170

Critics accuse Karen Hao of overstating AI data center water consumption in Empire of AI by up to five orders of magnitude due to unit conversion errors and misrepresented figures — Hao issued a correction after confirming a unit error.

19h|214.9k3.1k400

AM#1735|@ANDYMASLEY

PS#1482|@PETESKOMOROCH

AK#885|@ANDREY_KURENKOV

180

Antigravity releases UI and connectivity updates to version 2.0 that highlight the Open IDE button and add instructions for connecting to the IDE on Windows machines — New releases fix Windows IDE issues and reset weekly usage limits.

11h|229k3.1k288

VM#1144|@_MOHANSOLO

PS#934|@_PHILSCHMID

👩‍💻P#271|@DYNAMICWEBPAIGE

LK#90|@OFFICIALLOGANK

DH#5|@DEMISHASSABIS

190

Google tests experimental Gemma 4 app running fully offline on Pixel hardware with prototype glasses in remote wilderness without connectivity — Google Gemma and DeepMind engineers shared the demonstration online.

13h|66.7k1.1k320

𝚟⟢#1753|@VIEMCCOY

OB#1188|@OLIVIERBACHEM

HB#1022|@HERBIEBRADLEY

OS#490|@OSANSEVIERO

👩‍💻P#271|@DYNAMICWEBPAIGE

200

xAI Launches Overhauled Website Showcasing Models and Products

13h|2.1M6.3k838

EM#76|@ELONMUSK

210

DHH says GPT-5.5 shows substantial gains on complicated agent work, reversing version 5.2's lag behind Opus and marking a recovery for OpenAI amid rising competition — Jason Liu quote-tweeted the post and Adam.GPT reposted it.

11h|279.9k3.8k239

V(#1853|@REACH_VB

KA#488|@YACINEMTB

220

Google DeepMind updates Project Genie to integrate directly with Google Maps Street View for generating interactive 3D environments from real U.S. locations — Prototype at labs.google/fx/projectgenie outputs scenes labeled Genie 3.

3d|148k1k296

SF#1869|@SHLOMIFRUCHTER

RP#1032|@ROHANPAUL_AI

SL#1029|@SHUANGL13799063

JP#619|@JPARKERHOLDER

👩‍💻P#271|@DYNAMICWEBPAIGE

230

Dan Shipper's analysis of Every finds the company automated nearly all feasible tasks with AI agents yet grew its human workforce from 4 to 30 employees since GPT-3 — Report projects stronger effects as AI systems advance.

1d|118.4k358470

DS#1410|@DANSHIPPER

RP#1032|@ROHANPAUL_AI

240

Gemini Omni performed native video edits on the 1896 Lumiere film of a steam train, converting it into versions with a bullet train, LEGO styling, Muppets, and a time traveler while preserving reflections. — Wharton associate professor Ethan Mollick shared the 56-second example.

13h|159.9k2.1k605

FO#1996|@FOFRAI

NR#1562|@NATANIELRUIZG

👩‍💻P#271|@DYNAMICWEBPAIGE

EM#175|@EMOLLICK

250

OpenAI enabled a locked use feature for Codex that permits the system to control Mac applications from a connected phone even when the device is locked and the screen is off — It requires an Apple authorization plug-in for macOS sessions.

1d|691.7k8.2k1k

RH#804|@ROMAINHUET

260

Google's AI Overview responds to single-word verb queries such as 'disregard' and 'ignore' with command acknowledgments like 'Message received! Consider it forgotten — The pattern appeared in multiple X posts and a TechCrunch report.

14h|967.5k13.9k1.3k

PM#713|@PMINERVINI

DR#103|@DELIPRAO

270

Google's Gemini Omni model generates first-person video of a taxi following a route from a marked Google Maps screenshot with consistent viewpoint and visual details — Model synthesized dynamic simulation from static map imagery.

1d|127.5k1.1k208

SA#930|@SAMUELALBANIE

RG#538|@RUIQIGAO

👩‍💻P#271|@DYNAMICWEBPAIGE

AV#230|@AVDNOORD

LB#55|@GIFFMANA

280

Grok Build, an AI tool for web development, entered early beta with updates improving aesthetics and functionality plus plans for a recursive self-improvement loop — Boyuan Zheng posted the announcement, which Elon Musk quote-tweeted.

9h|#1VIEWS15.9M19.5k729

YS#383|@YSU_NLP

EM#76|@ELONMUSK

290

Researchers dispute Reflection AI's positioning as America's open source AI lab citing its empty model repository with zero public releases — Repository shows no models datasets or code after years of operation.

15h|65.3k1.1k56

CR#1444|@CWOLFERESEARCH

JS#1251|@JSUAREZ

VI#1195|@VIKHYATK

CB#999|@CODE_STAR

AA#436|@ARMENAGHA

NL#64|@NATOLAMBERT

MB#20|@MILES_BRUNDAGE

300

ChatGPT now supports uploading images of blank forms and completing them through voice mode dictation in the mobile app — Demo filled out a fitness membership registration form with spoken details.

11h|1.6M5.4k2.2k