Digg - AI news, before it trends

Today's Highlights

In case you missed it

OpenAI's internal reasoning model refutes the 1946 Erdős planar unit distance conjecture

AK#228|@BLACKHC

GM#153|@GARYMARCUS

SD#94|@_SHOLTODOUGLAS

RA#83|@_AROHAN_

SB#51|@SEBASTIENBUBECK

#1 BOOKMARKED

929

Sakana AI releases DiffusionBlocks to train neural networks one block at a time, cutting training memory up to 8x

FH#465|@FURONGH

HP#455|@HYHIEU226

HA#18|@HARDMARU

#1 VIEWED

Trajectory, co-founded by Ronak Malde, launches with $15M to build a continual learning platform for agentic AI models

RP#1032|@ROHANPAUL_AI

RS#321|@SCOBLEIZER

NL#64|@NATOLAMBERT

#1 LIKED

3.2k

Thibault Sottiaux, OpenAI Codex engineering lead, says GPT-5.2 and GPT-5.3-Codex will be sunset from Codex on June 2

RO#1220|@KRISHNANROHIT

TI#768|@THSOTTIAUX

Top Stories

Posts:—|Clusters:—|Next Crawl:—

10

Cognition raises over $1 billion at a $26 billion valuation, reaching a $492 million annual revenue run-rateThe Devin creator's enterprise usage grew tenfold this year.

1h|152.1k1.4k205

BH#1941|@BENHYLAK

WA#1906|@WALDEN_YAN

VV#1557|@VIPULVED

JW#1539|@WOLFEJOSH

BI#1402|@BILALTWOVEC

YP#1252|@YPATIL125

RP#1032|@ROHANPAUL_AI

SW#704|@SCOTTWU46

20

ESM project founder Alex Rives releases ESMFold2, claiming it beats AlphaFold3 on protein-protein interaction benchmarksLab tests validated binders targeting five therapeutic proteins.

5h|158.9k1.3k450

AP#1703|@ALEX_PEYS

AK#1675|@ANSHULKUNDAJE

BW#1661|@BOWANG87

AR#1087|@ALEXRIVES

30

Sakana AI releases DiffusionBlocks to train neural networks one block at a time, cutting training memory up to 8xIt treats the network's forward pass as a diffusion-like denoising process.

2h|127k1.4k#1BOOKMARKED929

FH#465|@FURONGH

HP#455|@HYHIEU226

HA#18|@HARDMARU

40

Trajectory, co-founded by Ronak Malde, launches with $15M to build a continual learning platform for agentic AI modelsIt uses real-time user corrections to post-train deployed models.

1h|#1VIEWS260.4k529324

RP#1032|@ROHANPAUL_AI

RS#321|@SCOBLEIZER

NL#64|@NATOLAMBERT

50

Spark Capital's Yasmin Razavi joins the Forbes Midas List after her $75 million Anthropic investment returned $3 billionShe backed Anthropic when the startup had no revenue

3h|50.1k34784

YR#1602|@YASMINRAZAVI

IZ#1273|@IVANHZHAO

FR#1200|@FRASER

New github stars (48hrs)

self-evolving/repo251d

Invokes AI agents via GitHub mentions to implement issues, review PRs, fix branches, and persist memory across runs.

modaic-ai/gepa-viz2591d

Renders GEPA prompt-optimization candidate trees as interactive force-directed graphs with live updates from a Python callback.

60

Amplify Partners GP Sarah Catanzaro argues file systems are better suited than databases for AI agent workloadsCurrent distributed file systems lack the concurrency agents require

1h|7.8k7566

SC#733|@SARAHCAT21

SS#407|@SH_REYA

AD#172|@ALEXGDIMAKIS

LB#55|@GIFFMANA

70

Thibault Sottiaux, OpenAI Codex engineering lead, says GPT-5.2 and GPT-5.3-Codex will be sunset from Codex on June 2Both models will remain accessible to developers via API

2h|189.3k#1LIKED3.2k187

RO#1220|@KRISHNANROHIT

TI#768|@THSOTTIAUX

80

OpenAI Foundation launches $250 million initiative to study and support workforce transitions in the age of AIThe program funds impact measurement and shared prosperity research.

33m|88.2k1.6k256#1COMMENTS522

WZ#49|@WOJ_ZAREMBA

90

ETH Zürich's Tiberiu Mușat proves that fixed-precision neural network weight norm is equivalent to Kolmogorov complexityThe equivalence explains why deep networks generalize rather than memorize.

4h|57.2k684600

LA#1822|@LUCAAMB

TC#120|@TACOCOHEN

100

Apollo CEO Marc Rowan argues AI's massive capital demands will outgrow venture equity and rely on private creditApollo manages more than $1 trillion in assets.

2h|37k129105

SW#1814|@SARAHDINGWANG

ET#1253|@ERIKTORENBERG

WM#1039|@WILLMANIDIS

110

Google's Logan Kilpatrick argues genuine understanding cannot be delegated to AI, prompting debate with researcher Andreas KirschKirsch argues continuous LLM interaction functionally replaces traditional comprehension.

2h|37.7k1k122

RISING LIKES1k

AK#228|@BLACKHC

LK#90|@OFFICIALLOGANK

120

MIT CSAIL's Alex Zhang open-sources a sandbox-free recursive language model training harness built on prime-rlThe companion 30B model is hosted on Hugging Face.

3h|37.1k604331

AZ#853|@A1ZHANG

OK#160|@LATEINTERACTION

130

Roboflow integrates RF-DETR into Hugging Face Transformers, claiming it outperforms YOLO on detection and segmentationThe integration supports low-VRAM fine-tuning on constrained hardware.

2h|31.6k439352

ME#675|@MERVENOYANN

C🤗#68|@CLEMENTDELANGUE

140

OpenAI's Roon argues aggregate token consumption is a highly valuable metric for tracking AI progressHe compared token metrics to pre-industrial steam power usage.

8h|30.1k52856

🍓🍓#1711|@IRULETHEWORLDMO

150

Pavel Izmailov and Andrew Gordon Wilson find LLMs can prevent catastrophic forgetting using self-generated replay dataThe method kept task evaluation accuracies near 100%.

2h|19k366236

PI#376|@PAVEL_IZMAILOV

AG#154|@ANDREWGWILS

160

RepoPrompt founder Eric Provencher joins OpenAI under Romain Huet after building the AI coding agentOpenAI is offering Codex credits to existing RepoPrompt users

2h|63.3k79263

RH#804|@ROMAINHUET

170

Hugging Face's Sayak Paul, Hila Chefer, and Linoy Tsaban announce an ECCV 2026 tutorial on post-training diffusion modelsThe sessions focus on model capability, control, and alignment.

6h|14.7k196107

KD#908|@CSPROFKGD

SD#80|@SEDIELEM

180

Wharton's Ethan Mollick says AI API tokens transitioned from an overlooked expense to an essential development requirement in one yearOpenClaw's creator reportedly spends $1 million monthly on tokens.

3h|32.4k40750

ND#1894|@NICKADOBOS

EM#175|@EMOLLICK

190

Tim Rocktäschel argues Google DeepMind's Pushmeet Kohli misses the potential for scientific autonomy by comparing AI to a telescopeKohli argues AI shifts scientists toward framing questions.

5h|9.4k15583

PK#298|@PUSHMEET

200

OpenAI's Boris Power builds a self-improving tax agent that reached 90% completion on nearly 60% of tasksThe Codex-powered agent learns directly from human corrections.

2h|5.7k5222

RP#1032|@ROHANPAUL_AI

BP#357|@BORISMPOWER

Rising Stories

AI Tools2H AGO
Teaser surfaces for Codex on Windows
#1496@KIMMONISMUS
#687@TUNGUZ
656.7k
AI Education1H AGO
UC STEM professors urge return to standardized testing
#1670@KIMMAICUTLER
#270@GARRYTAN
60569.5k
AI Infrastructure1H AGO
ClickHouse hits $250M ARR after 10 years open source
#1511@MATTTURCK
#327@VKHOSLA
4511.9k
AI Infrastructure47M AGO
Redpoint releases 2026 InfraRed report on agent runtimes
#339@WILLCCBB
#64@NATOLAMBERT
427.7k
AI Costs3H AGO
Ethan Mollick notes AI token spend now essential for devs
#1894@NICKADOBOS
#175@EMOLLICK
40732.4k

Github Stars

(7 days)

manaflow-ai/cmux2D AGO AGO
Integrates Ghostty into a macOS terminal with vertical tabs, sidebar metadata, OSC notifications, and in-app browser for AI coding agents.
759
660
30
495
19.8k stars
self-evolving/repo1D AGO AGO
Invokes AI agents via GitHub mentions to implement issues, review PRs, fix branches, and persist memory across runs.
594
878
25 stars
modaic-ai/gepa-viz1D AGO AGO
Renders GEPA prompt-optimization candidate trees as interactive force-directed graphs with live updates from a Python callback.
878
793
259 stars

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

Yesterday's Top Stories, May 26, 2026.

Posts: 3,375|Clusters: 1,001|Stories frozen at midnight PT

10

OpenAI's internal reasoning model refutes the 1946 Erdős planar unit distance conjectureNicknamed Mythos, the model used algebraic field methods.

6d|1.4M8.7k1.4k

CH#1496|@KIMMONISMUS

BI#1402|@BILALTWOVEC

DA#1271|@DAMEKDAVIS

AV#1245|@ALTRYNE

AS#1054|@ANDERSSANDBERG

LA#980|@SCALING01

AD#867|@DORIALEXANDER

EL#716|@ELIEBAKOUCH

HE#572|@HEINRICHKUTTLER

AC#517|@ANDREWCURRAN_

CL#442|@CHENGLEISI

AK#228|@BLACKHC

GM#153|@GARYMARCUS

SD#94|@_SHOLTODOUGLAS

RA#83|@_AROHAN_

SB#51|@SEBASTIENBUBECK

20

Serena Ge releases DeepSWE, a long-horizon benchmark designed to evaluate AI coding agents on complex engineering tasksIts prompts are half the length of SWE-bench Pro.

14h|1.4M10.7k3.2k

ND#1894|@NICKADOBOS

🍓🍓#1711|@IRULETHEWORLDMO

CH#1496|@KIMMONISMUS

FB#1153|@XEOPHON

B(#982|@ANDERSONBCDEFG

LA#980|@SCALING01

TS#765|@SANDERSTED

T(#420|@TEORTAXESTEX

GT#270|@GARRYTAN

LB#55|@GIFFMANA

30

Yo Shavit, OpenAI's frontier AI safety policy lead, resigns to join the OpenAI Foundation's AI Resilience programHe cited a limited timeline to prepare for superintelligence

11h|217k1.6k248

JT#1871|@JACOBTREF

HB#1459|@HAYDNBELFIELD

SÓ#1457|@S_OHEIGEARTAIGH

NC#1319|@CHOWDHURYNEIL

AV#1245|@ALTRYNE

MW#1241|@MILESKWANG

RO#1220|@KRISHNANROHIT

HB#1022|@HERBIEBRADLEY

🎭#878|@DEEPFATES

YS#638|@YONASHAV

LG#290|@NABLA_THETA

NL#64|@NATOLAMBERT

SB#27|@SLEEPINYOURHAT

MB#20|@MILES_BRUNDAGE

40

China reportedly restricts overseas travel for AI talent at Alibaba and DeepSeek to protect algorithmic secretsCritics note DeepSeek already publishes its AI models openly.

22h|394.5k5.7k783

ZE#1497|@ZEPHYR_Z9

CY#1301|@CHHAVIYADAV_

RO#1220|@KRISHNANROHIT

RP#1032|@ROHANPAUL_AI

KA#488|@YACINEMTB

T(#420|@TEORTAXESTEX

FI#397|@FINBARRTIMBERS

YM#368|@YIMATWEETS

NL#64|@NATOLAMBERT

50

Pope Leo XIV releases AI ethics encyclical Magnifica Humanitas as Catholic Church partners with AnthropicAnthropic co-founder Chris Olah is scheduled to speak.

8d|344k2.4k673

AL#1114|@ALTH0U

NL#1006|@LABENZ

JA#837|@APPLES_JIMMY

PD#654|@PMDDOMINGOS

DW#392|@DEANWBALL

TM#359|@ISCIENCELUVR

BC#322|@BCHERNY

MB#20|@MILES_BRUNDAGE

60

Elon Musk rejects commentator's claim that xAI cannot reach the top industry position within three yearsMusk noted he co-founded OpenAI ten years ago.

17h|#1VIEWS5.8M#1LIKED70.1k3k

B(#839|@BEFFJEZOS

KA#488|@YACINEMTB

T(#420|@TEORTAXESTEX

BY#341|@BILLYUCHENLIN

EM#76|@ELONMUSK

70

Pope Leo XIV proposes classifying patents, algorithms, and data as universal goods to combat digital exclusionTwitter co-founder Jack Dorsey publicly endorsed the proposal.

19h|1.6M35.7k2.8k

C🎉#863|@CHARLES_IRL

VW#707|@VINCENTWEISSER

TM#359|@ISCIENCELUVR

WB#339|@WILLCCBB

80

MiniMax details its M3 sparse attention architecture, claiming a 15.6x decoding speedup at 1 million tokensThe design also achieves a 9.7x prefill speedup.

15h|690.8k4.7k1.3k

ZE#1497|@ZEPHYR_Z9

CH#1496|@KIMMONISMUS

LA#980|@SCALING01

EL#716|@ELIEBAKOUCH

PM#713|@PMINERVINI

T(#420|@TEORTAXESTEX

90

Creator Theo and founder DHH say GPT-5.5 excels at software engineering when guided by high-level steering and custom agent filesTheo spent two months tuning his prompts and agents.md files.

1d|303.7k4.2k564

EZ#1965|@ERICZAKARIASSON

🍓🍓#1711|@IRULETHEWORLDMO

RH#804|@ROMAINHUET

PS#495|@STEIPETE

100

Uber COO Andrew Macdonald says rising AI token costs are hard to justify after exhausting its 2026 Claude Code budgetUber is slowing down hiring to offset the ballooning costs

1d|116.9k567121

AI#1777|@ALEXOLEGIMAS

CE#1598|@INTUITMACHINE

C🎉#863|@CHARLES_IRL

AJ#486|@JM_ALEXIA

GM#153|@GARYMARCUS

FC#37|@FCHOLLET

110

New York intelligence assessment warns AI could spark "anti-tech violent extremism" targeting data centersThe designation may trigger FBI probes of AI safety groups

19h|880.2k10k2.1k

GL#1984|@GARRISONLOVELY

PR#1332|@8TEAPI

AC#517|@ANDREWCURRAN_

120

OpenRouter raises $113 million Series B at a $1.3 billion valuation as weekly traffic hits 25 trillion tokensCapitalG led the funding for the multi-model routing service.

16h|329.3k2.8k370

GK#1182|@GREGKAMRADT

CB#999|@CODE_STAR

DE#374|@DEEDYDAS

130

Josef Chen releases EPICURE, a 2MB factor decomposition model mapping 1,790 ingredients from 4.1 million recipesIt identified five emergent culinary clusters without prior labels.

11h|248.9k3.8k3k

YO#1234|@YOHEINAKAJIMA

RS#612|@ZIV_RAVID

AC#263|@ANDREW_N_CARR

140

Guri Singh argues LLM progress validates the 'Stochastic Parrots' paper, while creator Jon Stokes claims it has aged poorlyThe 2020 paper warned of hallucinations and systemic bias.

18h|615.7k8.7k#1BOOKMARKED4.9k

PS#1482|@PETESKOMOROCH

S(#1303|@SGUADA

150

Xiaomi permanently cuts MiMo-V2.5 and MiMo-V2.5-Pro API prices by up to 99 percent on cache hitsCritics argue the deep discounts rely on financial subsidies.

14h|647.8k4k1.6k

CH#1496|@KIMMONISMUS

JA#837|@APPLES_JIMMY

T(#420|@TEORTAXESTEX

160

PrismML releases Bonsai Image 4B, a 930 MB 1-bit diffusion transformer that generates smartphone images in nine secondsThe Apache 2.0 model is based on FLUX.2 Klein.

12h|129.6k1.4k826

LN#1709|@CLASHLUKE

SR#604|@CLONEOFSIMO

RA#83|@_AROHAN_

170

Open-source creator Jason warns that generating HTML via Claude Code spends up to 40% of tokens on structural markupAnthropic's Thariq Shihipar originally shared the non-technical workflow.

10h|265.2k2.9k1.4k

EZ#1965|@ERICZAKARIASSON

MB#1759|@MATTHEWBERMAN

180

Sangyun Lee and Giulia Fanti propose a "sleep" phase to convert LLM context into fast weights and clear KV cacheOffline consolidation passes reduce quadratic attention costs during inference.

20h|96.4k1.4k1k

MN#1179|@MENHGUIN

AK#885|@ANDREY_KURENKOV

EL#475|@OMARSAR0

TM#359|@ISCIENCELUVR

190

Tony Zador finds only Claude Opus and GPT-5.5 recognize a Star Trek parody satirizing AI failure justificationsSonnet, Haiku, and Gemini failed to grasp the satire.

14h|927.6k38.7k2.6k

TZ#1925|@TONYZADOR

XL#1674|@XLR8HARDER

200

Aggressive Forza Horizon player Bowie Knife99 poisons game's Drivatar AI training data, creating clone bots that ram other driversThe incident demonstrates unintentional training data poisoning in gaming

8h|697.9k9k1.9k

ZN#537|@ZACHARYNADO

T(#420|@TEORTAXESTEX

210

Princeton University's NLP group graduates 11 PhD students, including the first doctoral cohort of Danqi ChenThe cohort focused on language model development and analysis.

14h|54.5k58756

CY#1301|@CHHAVIYADAV_

MX#718|@XIAMENGZHOU

ZL#703|@LIUZHUANG1234

KN#599|@KARTHIK_R_N

DC#136|@DANQI_CHEN

OP#72|@OFIRPRESS

220

Bilawal Sidhu, former Google PM, demonstrates generating motion-controlled POV video from hand-drawn path inputs using the Omni AI modelSidhu suggests incorporating spatial RAG to improve geographic accuracy.

2d|195.5k943492

BS#1909|@BILAWALSIDHU

ND#1894|@NICKADOBOS

NR#1562|@NATANIELRUIZG

SM#530|@SWAROOPRM7

230

Merge API launches Merge Gateway to help enterprises build custom model routers based on product-specific benchmarksIt prevents enterprise cost overruns from overusing frontier models

16h|3.5M2.3k1.4k

CE#1598|@INTUITMACHINE

CH#1496|@KIMMONISMUS

240

Nathan Lambert of the Allen Institute for AI says Google's Gemma 4 models are outpacing Alibaba's Qwen in cumulative downloadsAnalysts cite Google's Apache 2.0 license as the driver

14h|169.7k719161

FB#1153|@XEOPHON

👩‍💻P#271|@DYNAMICWEBPAIGE

NL#64|@NATOLAMBERT

MB#20|@MILES_BRUNDAGE

250

AngelList's Naval Ravikant argues the primary AI competition is humans with AI versus humans withoutResearch engineer kache questioned the future of developers who fully automate.

17h|399.7k12k1.3k

KA#488|@YACINEMTB

260

Nous Research co-founder Teknium says xAI improved caching and reset usage limits on Grok Build BetaThe reset follows reports of users rapidly hitting caps.

9h|623.9k5.9k196

BY#341|@BILLYUCHENLIN

T🪽#259|@TEKNIUM

EM#76|@ELONMUSK

270

Systems engineer Yacine recommends training models on a single GPU in under one minute to maximize research learning ratesLucas Beyer clarified BiternionNet was developed at RWTH Aachen

16h|203.8k3.1k2.6k

KA#488|@YACINEMTB

SD#80|@SEDIELEM

LB#55|@GIFFMANA

280

Figure AI founder Brett Adcock announces deal to deploy humanoid robots across Catalyst Brands' logistics networkInitial Figure F03 deployments will begin in Reno, Nevada.

11h|782.4k6.2k828

BA#688|@ADCOCK_BRETT

290

Epoch AI's Jaime Sevilla and Luke Emberson warn that surging token demand will outpace global Blackwell capacity through 2032The shortage could force developers to deploy smaller models.

10h|25.1k275136

SH#1488|@HAMANDCHEESE

JS#1389|@JSEVILLAMOL

CP#1092|@CHRISPAINTERYUP

HB#1022|@HERBIEBRADLEY

SK#505|@SEBKRIER

300

Poolside AI releases technical report for Laguna M.1 and XS.2 models detailing its Model Factory infrastructureThe report details continuous infrastructure over static model cards

15h|29.4k281193

EL#716|@ELIEBAKOUCH

AC#263|@ANDREW_N_CARR

AK#28|@_AKHALIQ

SH#14|@SARAHOOKR