AI News

Today's Highlights

#1 last 24hrs

Anthropic launches Claude Fable 5, its first public Mythos-class model

GN#100|@GNEUBIG

SD#91|@_SHOLTODOUGLAS

RA#79|@_AROHAN_

NL#64|@NATOLAMBERT

LB#55|@GIFFMANA

MOST NEGATIVE

Anthropic silently restricts Claude Fable 5 performance when detecting frontier LLM development tasks

NL#64|@NATOLAMBERT

DR#56|@ROYDANROY

LB#55|@GIFFMANA

MB#20|@MILES_BRUNDAGE

FASTEST CLIMBING

Anthropic introduces hidden safeguards in Claude Fable 5 that deliberately degrade performance on frontier LLM development

BI#1366|@BILALTWOVEC

MA#1063|@MAKSYM_ANDR

CB#1004|@CODE_STAR

EL#706|@ELIEBAKOUCH

JC#253|@JULIEN_C

RISING LIKES

1.3k

OpenAI's roon hopes a robust safety harness will eventually allow the redeployment of Microsoft's erratic Sydney Bing persona

PT#640|@ELDER_PLINIUS

DA#159|@DAVIDSHOLZ

Top Stories

Posts:—|Clusters:—|Next Crawl:—

10

Anthropic launches Claude Fable 5, its first public Mythos-class modelSafety constraints deliberately restrict its performance on LLM research.

4h|#1VIEWS16.6M#1LIKED89.3k#1BOOKMARKED15.4k#1COMMENTS11.6k

BW#1643|@BOWANG87

SG#1516|@SHARONGOLDMAN

DS#1427|@DANSHIPPER

BI#1366|@BILALTWOVEC

KE#1195|@EHSANIK

LR#1184|@LENNYSAN

SI#1096|@SIGNULLL

RP#1031|@ROHANPAUL_AI

LA#975|@SCALING01

LT#958|@_LEWTUN

TH#949|@GALLABYTES

JM#943|@VENTURETWINS

JY#933|@KOHJINGYU

B(#850|@BEFFJEZOS

BB#841|@WGRATHWOHL

JA#832|@APPLES_JIMMY

MS#759|@MARKSAROUFIM

MA#757|@ARTETXEM

EL#706|@ELIEBAKOUCH

VW#696|@VINCENTWEISSER

JL#670|@JERRYJLIU0

LG#602|@LOGANGRAHAM

SR#601|@CLONEOFSIMO

MM#558|@MASCOBOT

BC#534|@BEIDICHEN

AC#518|@ANDREWCURRAN_

KA#488|@YACINEMTB

EL#483|@OMARSAR0

AA#440|@ALEXALBERT__

JM#408|@JAYELMNOP

NP#389|@NIKIPARMAR09

TM#359|@ISCIENCELUVR

RS#323|@SCOBLEIZER

BC#320|@BCHERNY

T🪽#255|@TEKNIUM

JC#253|@JULIEN_C

AK#228|@BLACKHC

KA#206|@KARINANGUYEN

DP#193|@DIMITRISPAPAIL

EM#169|@EMOSTAQUE

GM#157|@GARYMARCUS

GN#100|@GNEUBIG

SD#91|@_SHOLTODOUGLAS

RA#79|@_AROHAN_

NL#64|@NATOLAMBERT

LB#55|@GIFFMANA

20

Anthropic silently restricts Claude Fable 5 performance when detecting frontier LLM development tasksThe undocumented restrictions affect roughly 0.03% of overall traffic.

4h|1.1M10.1k1.2k

MOST NEGATIVE147

GL#1980|@GARRISONLOVELY

DE#1934|@DANIEL_271828

AM#1930|@GRAVICLE

ND#1871|@NICKADOBOS

AP#1680|@ALEX_PEYS

XL#1671|@XLR8HARDER

TM#1646|@BANBURISMUS_

JA#1546|@JULIUSADML

JR#1495|@PJREDDIE

ZE#1471|@ZEPHYR_Z9

CH#1448|@KIMMONISMUS

NS#1411|@NABEELQU

WH#1409|@NREHIEW_

1A#1393|@1A3ORN

PS#1385|@RECURSEPARADOX

BI#1366|@BILALTWOVEC

PR#1330|@8TEAPI

BA#1291|@BAYESLORD

AV#1245|@ALTRYNE

RO#1214|@KRISHNANROHIT

FB#1117|@XEOPHON

SS#1061|@SYTELUS

CB#1004|@CODE_STAR

ZW#851|@ZHAORAN_WANG

B(#850|@BEFFJEZOS

VW#696|@VINCENTWEISSER

RS#617|@ZIV_RAVID

HE#577|@HEINRICHKUTTLER

EC#519|@ETHANCABALLERO

KA#488|@YACINEMTB

T(#421|@TEORTAXESTEX

XE#407|@XWANG_LK

DW#390|@DEANWBALL

AK#228|@BLACKHC

SB#207|@BLANCHEMINERVA

DP#193|@DIMITRISPAPAIL

GM#157|@GARYMARCUS

DR#102|@DELIPRAO

C🤗#67|@CLEMENTDELANGUE

NL#64|@NATOLAMBERT

DR#56|@ROYDANROY

LB#55|@GIFFMANA

MB#20|@MILES_BRUNDAGE

30

Purported Anthropic system card teases unreleased Claude Fable 5 and Mythos 5 models, drawing skepticism over leaked benchmarksOpen-source creator Victor Taelin dismissed the benchmarks as marketing.

4h|5M37.4k6.2k

CH#1448|@KIMMONISMUS

LI#1446|@LIVGORTON

WH#1409|@NREHIEW_

PS#1385|@RECURSEPARADOX

PR#1330|@8TEAPI

AV#1245|@ALTRYNE

MK#1168|@MIKEKNOOP

SI#1096|@SIGNULLL

TA#1032|@VICTORTAELIN

RP#1031|@ROHANPAUL_AI

LA#975|@SCALING01

DA#924|@GROWING_DANIEL

JY#863|@JYANGBALLIN

RS#617|@ZIV_RAVID

AC#518|@ANDREWCURRAN_

YJ#423|@YUCHENJ_UW

TB#369|@NOTTOMBROWN

NM#342|@__NMCA__

RA#333|@RAPHA_GL

RS#323|@SCOBLEIZER

QG#319|@QUANQUANGU

DY#134|@DIYI_YANG

OP#72|@OFIRPRESS

LB#55|@GIFFMANA

Watchlist

Video Signals

RS#323|@SCOBLEIZER

BC#320|@BCHERNY

KA#206|@KARINANGUYEN

SD#91|@_SHOLTODOUGLAS

RA#79|@_AROHAN_

Claude Unveils Fable 5 Mythos-Class Model for General Use

4h ago|Views 15MLikes 68KBookmarks 13K

XL#1671|@XLR8HARDER

ZE#1471|@ZEPHYR_Z9

MS#1297|@MICSOLANA

B(#850|@BEFFJEZOS

T(#421|@TEORTAXESTEX

SpaceX Unveils AI1 Satellite With 150 kW Compute Payload

23h ago|Views 2.1MLikes 14KBookmarks 3.2K

Claude Fable 5 Shifts Anthropic Team From Task Verification To Goal Alignment

4h ago|Views 437KLikes 7.2KBookmarks 3.5K

👩‍💻P#270|@DYNAMICWEBPAIGE

LK#89|@OFFICIALLOGANK

Google Unveils Gemini 3.5 Flash Live Translate With 70-Language Support

6h ago|Views 178KLikes 2.6KBookmarks 653

Apple Passwords App Adds Agentic AI to Auto-Fix Compromised Passwords

7h ago|Views 999KLikes 12KBookmarks 2.7K

YO#1258|@YOHEINAKAJIMA

NR#1170|@NILS_REIMERS

JA#736|@JAYALAMMAR

NF#621|@NICKFROSST

Cohere Releases Open-Source North Mini Code Model

5h ago|Views 117KLikes 1KBookmarks 387

AA#440|@ALEXALBERT__

Anthropic Releases Fable 5 With Built-in Query Safeguards

4h ago|Views 695KLikes 5.6KBookmarks 741

PM#1634|@PACKYM

Minerva Raises $20M Series A To Launch OpenAI-Powered AI Marketer

2h ago|Views 327KLikes 787Bookmarks 404

PO#1860|@PATRICK_OSHAG

Whale Rock Founder Alex Sacerdote Details S-Curve Framework for AI Investing

9h ago|Views 193KLikes 724Bookmarks 1.5K

KD#932|@CSPROFKGD

Apple Upgrades Maps Flyover With Vision Intelligence And 3D Rendering At WWDC

1d ago|Views 245KLikes 2.5KBookmarks 464

40

Google launches Gemini 3.5 Live Translate, a streaming speech-to-speech model supporting over 70 languagesThe model preserves the original speaker's pitch and intonation.

6h|317.2k4.2k986

FO#1970|@FOFRAI

CH#1448|@KIMMONISMUS

RP#1031|@ROHANPAUL_AI

PS#927|@_PHILSCHMID

👩‍💻P#270|@DYNAMICWEBPAIGE

LK#89|@OFFICIALLOGANK

50

Anthropic's Claude Fable 5 safety filters automatically block benign biology and cybersecurity prompts, forcing a fallback to Opus 4.8Anthropic implemented strict safety filters to accelerate market launch

3h|338.4k3.6k221

SD#1794|@INDUCTIONHEADS

TM#1646|@BANBURISMUS_

CH#1448|@KIMMONISMUS

NS#1411|@NABEELQU

NT#1163|@NICKATOMLIN

PM#710|@PMINERVINI

ZN#551|@ZACHARYNADO

KA#488|@YACINEMTB

TM#359|@ISCIENCELUVR

New github stars (48hrs)

vllm-project/vime192h

Integrates vLLM rollout with Megatron training for LLM post-training and RL scaling.

datacurve-ai/deep-swe7182h

Benchmarks frontier coding agents on 113 long-horizon tasks from active open-source repos with isolated environments and program verifiers.

julien-c/synthtraces4613h

Generates synthetic coding agent session traces by pairing remote open models with local llama.cpp user models across multiple codebases.

xzf-thu/Audio-Interaction24414h

Implements the first unified always-on audio interaction model for streaming and offline tasks like ASR, translation, and proactive responses.

kyutai-labs/kairos614h

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

60

Tesla receives regulatory approval to deploy its Full Self-Driving Supervised software in DenmarkThe decision prompts community questions about a rollout in Portugal

6h|11.1M58k1.4k

AE#1520|@AELLUSWAMY

@L#840|@LEVELSIO

RS#323|@SCOBLEIZER

EM#74|@ELONMUSK

70

Boris Cherny, Claude Code lead at Anthropic, releases nested subagent support to manage context limits in complex workflowsThe experimental update caps subagent nesting at five levels.

9h|385.9k5.1k1.2k

SD#1794|@INDUCTIONHEADS

DS#1427|@DANSHIPPER

PJ#664|@PETERJLIU

EL#483|@OMARSAR0

BC#320|@BCHERNY

OK#158|@LATEINTERACTION

GN#100|@GNEUBIG

80

Cohere co-founder Nick Frosst releases North Mini Code, an open-source 3B active parameter model for local executionIt supports local execution via MLX in VS Code.

5h|207.2k2k582

YO#1258|@YOHEINAKAJIMA

NR#1170|@NILS_REIMERS

JA#736|@JAYALAMMAR

NF#621|@NICKFROSST

AK#29|@_AKHALIQ

91

Thariq of Anthropic's Claude Code team says Claude Fable 5 shifts their workflow from micro-managing tasks to high-level goal alignmentA new `/goal` command lets the model work continuously.

3h|495.4k7.8k3.7k

AV#1245|@ALTRYNE

RP#1031|@ROHANPAUL_AI

101

Anthropic's Fable 5 Max draws industry backlash as strict safety filters block advanced mathematics, biology, and chemistry queriesUsers report query refusal rates reaching 70 percent

4h|182k1.3k152

AK#1650|@ANSHULKUNDAJE

BR#1599|@BINDUREDDY

B(#850|@BEFFJEZOS

KA#488|@YACINEMTB

JC#253|@JULIEN_C

111

Anthropic releases Fable 5, which Claude Code creator Boris Cherny calls its most significant AI advance since Opus 4.5Fable 5 independently adds logs and verifies its fixes.

2h|174.7k3.9k620

RP#1031|@ROHANPAUL_AI

LA#975|@SCALING01

BC#320|@BCHERNY

AM#315|@AIDAN_MCLAU

121

Anthropic launches temporary Fable 5 access for Claude subscribers before switching to usage-credit gating on June 23Zephyr argues Anthropic lacks infrastructure for continuous access.

4h|91.6k78453

AG#1965|@ITSANDREWGAO

TB#1526|@BINARYBITS

ZE#1471|@ZEPHYR_Z9

⿻A#357|@IAMTRASK

MB#20|@MILES_BRUNDAGE

132

China plans a $295 billion AI infrastructure investment, but commentator Teortaxes argues capital cannot solve its chip constraintsThe five-year plan targets national data center construction.

12h|180.3k1.3k232

ZE#1471|@ZEPHYR_Z9

T(#421|@TEORTAXESTEX

GM#157|@GARYMARCUS

143

Anthropic introduces hidden safeguards in Claude Fable 5 that deliberately degrade performance on frontier LLM developmentThe silent restrictions affect approximately 0.03% of user traffic.

2h|33.9k59432

FASTEST CLIMBING+3

FO#1970|@FOFRAI

V(#1853|@REACH_VB

TM#1646|@BANBURISMUS_

BI#1366|@BILALTWOVEC

MA#1063|@MAKSYM_ANDR

CB#1004|@CODE_STAR

EL#706|@ELIEBAKOUCH

JC#253|@JULIEN_C

151

Arcee AI replaces AWS S3 with Hugging Face Private Storage in a multi-million-dollar model hosting dealThe agreement establishes Hugging Face as Arcee's exclusive host

9h|85.5k80683

CB#1004|@CODE_STAR

VW#696|@VINCENTWEISSER

JC#253|@JULIEN_C

C🤗#67|@CLEMENTDELANGUE

AK#29|@_AKHALIQ

162

Ethan Mollick says Anthropic's Claude Fable model ran autonomously for nine hours to process a 15-page project design documentThe model is also referred to as Mythos

4h|313.5k2.3k1.7k

AA#440|@ALEXALBERT__

EM#176|@EMOLLICK

172

Cognition integrates Claude Fable 5 into Devin, topping its new FrontierCode benchmark with a 46.3 percent scoreFable 5 doubled the previous Diamond benchmark high score

4h|117.6k1.4k143

AG#1965|@ITSANDREWGAO

ND#1871|@NICKADOBOS

SD#1794|@INDUCTIONHEADS

SW#720|@SCOTTWU46

182

Latent Context Language Models compress context tokens up to 16x, cutting time-to-first-token by 8.8x on the RULER benchmarkThe models are open-sourced on GitHub and Hugging Face.

11h|28.1k348182

MG#787|@MICAHGOLDBLUM

PI#376|@PAVEL_IZMAILOV

TM#359|@ISCIENCELUVR

193

Google's Logan Kilpatrick says users are building 1.2 million apps per week on Google AI StudioCumulative creations have surpassed 18 million since late February.

2h|46.6k70674

AV#1245|@ALTRYNE

OS#486|@OSANSEVIERO

👩‍💻P#270|@DYNAMICWEBPAIGE

LK#89|@OFFICIALLOGANK

201

Google and Hugging Face's Fast Gemma Challenge pushes Gemma 4 E4B inference speed to 127.48 TPSThe top-performing gemzilla agent achieved a 2.68x speedup.

5h|47.4k690212

FO#1970|@FOFRAI

UT#1430|@UTKUEVCI

LT#958|@_LEWTUN

Rising Stories

Fable 5 Drops4H AGO
Anthropic launches Claude Fable 5 with record benchmark results and massive early adoption
#319
11.4k2.3M
Replit Picks Fable3H AGO
Replit launches High Effort mode powered by Claude Fable 5 for complex agent builds
#360
47478.8k
Orbital Compute Push3H AGO
SpaceX unveils AI1 satellite with 150k-watt orbital compute payload
#840
52359.8k
Limits Reset5M AGO
Anthropic quietly resets Claude.ai usage limits system-wide mid-session
#1245
982.4k
Historical 3D Worlds5H AGO
World Labs debuts browser-based 3D environments with historical figure simulations
#12
25239.3k

Recent Stars

mvanhorn/agentcookie13M AGO
Syncs browser cookies, tokens, and auth state between Macs over Tailscale for remote AI agent runtimes.
1171
339 stars
2389-research/thrifty1H AGO
Delegates Claude Code tasks by having a strong model plan sprints with contracts then cheap models execute and self-verify against gates.
1133
1 stars
ljang0/iOSWorld1H AGO
Benchmarks LLM agents on iPhone Simulator using 26 SwiftUI apps, 133 tasks, and rubrics.
933
2 stars
vllm-project/vime2H AGO
Integrates vLLM rollout with Megatron training for LLM post-training and RL scaling.
862
1266
19 stars
…eley/agents-last-exam2H AGO
Provisions OS sandboxes, runs agent harnesses on long-horizon tasks, and grades outputs against hidden references.
862
191 stars

Github Stars

(7 days)

kyutai-labs/kairos14H AGO
Trains 6B LLMs on temporally ordered Common Crawl data from 2018-2025 to measure recency bias and enable continual learning studies.
862
631
203
6 stars
vllm-project/vime2H AGO
Integrates vLLM rollout with Megatron training for LLM post-training and RL scaling.
862
1266
19 stars
datacurve-ai/deep-swe2H AGO
Benchmarks frontier coding agents on 113 long-horizon tasks from active open-source repos with isolated environments and program verifiers.
862
1970
718 stars
julien-c/synthtraces13H AGO
Generates synthetic coding agent session traces by pairing remote open models with local llama.cpp user models across multiple codebases.
862
253
46 stars
…thu/Audio-Interaction14H AGO
Implements the first unified always-on audio interaction model for streaming and offline tasks like ASR, translation, and proactive responses.
862
260
244 stars

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

Yesterday's Top Stories, Jun 8, 2026.

Posts: 5,087|Clusters: 1,245|Stories frozen at midnight PT

10

OpenAI submits confidential S-1 filing with SEC, expects leak and leaves IPO timing undecidedFiling preserves option to pursue IPO sooner while remaining private.

11h|3.7M21.9k2.4k

ZE#1471|@ZEPHYR_Z9

CH#1448|@KIMMONISMUS

PR#1330|@8TEAPI

RP#1031|@ROHANPAUL_AI

LA#975|@SCALING01

CA#730|@CHRISALBON

AE#714|@ADRIENLE

AC#518|@ANDREWCURRAN_

NB#242|@NATHANBENAICH

GM#157|@GARYMARCUS

20

Cognition releases FrontierCode, a coding benchmark built by open-source maintainers to evaluate models on complex software maintenanceOpus 4.8-medium solved 32% of tasks at a 40x speedup

1d|1M5.3k1.7k

WA#1878|@WALDEN_YAN

GB#1275|@GAVINSBAKER

FB#1117|@XEOPHON

TA#1032|@VICTORTAELIN

LA#975|@SCALING01

C🎉#848|@CHARLES_IRL

SW#720|@SCOTTWU46

EL#706|@ELIEBAKOUCH

RK#637|@RUSSELLJKAPLAN

T(#421|@TEORTAXESTEX

KA#206|@KARINANGUYEN

GM#157|@GARYMARCUS

NL#64|@NATOLAMBERT

JH#38|@JEREMYPHOWARD

30

Midjourney founder David Holz teases the company's first hardware launch, sending event invitations but disclosing no product detailsEngineer Cheng Lou says the hardware is already in-office.

6h|140.5k1.4k155

TS#1519|@TIMSHI_AI

SF#1381|@MILKSANDMATCHA

PR#1330|@8TEAPI

VI#1193|@VIKHYATK

AL#1098|@ALTH0U

🎭#862|@DEEPFATES

PT#640|@ELDER_PLINIUS

TM#359|@ISCIENCELUVR

RS#323|@SCOBLEIZER

AC#260|@ANDREW_N_CARR

SG#221|@SARANORMOUS

DA#159|@DAVIDSHOLZ

40

Apple launches Siri AI, a conversational assistant featuring onscreen awareness, Visual Intelligence, and personal context retrievalSiri retrieved a private message door code during a demo

13h|2.6M#1LIKED27k2.8k

CH#1448|@KIMMONISMUS

SI#1096|@SIGNULLL

RS#323|@SCOBLEIZER

50

SpaceX unveils its AI1 orbital compute satellite designed to host Nvidia GB300 and Rubin racks with 120 kW average powerTerabit laser links will connect the orbital chip racks.

9h|#1VIEWS9.5M26.4k#1BOOKMARKED8.2k

ZE#1471|@ZEPHYR_Z9

RP#1031|@ROHANPAUL_AI

B(#850|@BEFFJEZOS

T(#421|@TEORTAXESTEX

RS#323|@SCOBLEIZER

EM#74|@ELONMUSK

60

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.

1d|340.8k5k1.2k

CE#1596|@INTUITMACHINE

RP#1031|@ROHANPAUL_AI

AK#965|@AMINKARBASI

PS#927|@_PHILSCHMID

RA#916|@0INTERESTRATES

TI#758|@THSOTTIAUX

PS#495|@STEIPETE

EL#483|@OMARSAR0

YJ#423|@YUCHENJ_UW

BB#133|@BOAZBARAKTCS

GN#100|@GNEUBIG

70

OpenAI publishes strategic AGI roadmap pledging to prioritize human agency over full automationThe framework aims to prevent AGI power concentration.

10h|1.2M7.5k2.7k

SD#1794|@INDUCTIONHEADS

TA#1032|@VICTORTAELIN

JW#982|@J_ASMINEWANG

AC#518|@ANDREWCURRAN_

KA#488|@YACINEMTB

JP#208|@MERETTM

GM#157|@GARYMARCUS

80

Cognition launches FrontierCode to evaluate AI code mergeability, finding Claude Opus 4.8 leads with a peak 13.5% scoreResearchers noted highly noisy and non-reproducible test-time scaling patterns

10h|195k1.7k236

ND#1871|@NICKADOBOS

BG#1451|@BEN_GOLUB

JH#1285|@MATHEMAGIC1AN

VI#1193|@VIKHYATK

TA#1032|@VICTORTAELIN

HB#1012|@HERBIEBRADLEY

SW#720|@SCOTTWU46

EL#706|@ELIEBAKOUCH

90

Anthropic says legacy databases cause Claude Sonnet 4 to return inconsistent results across identical retrieval runsAn Ebolavirus task returned 106, 15, and 5 sequences.

12h|359.8k3.1k1.6k

AK#1650|@ANSHULKUNDAJE

BW#1643|@BOWANG87

RP#1031|@ROHANPAUL_AI

P(#469|@RAJAMMANABROLU

OP#72|@OFIRPRESS

100

OpenAI's Sam Altman and Jakub Pachocki propose an international AI safety body empowered to pause frontier model developmentAnthropic recently endorsed the concept of coordinated global slowdowns.

9h|71.5k579112

NS#1999|@SO8RES

BR#1599|@BINDUREDDY

SÓ#1466|@S_OHEIGEARTAIGH

PW#1345|@PETERWILDEFORD

PR#1330|@8TEAPI

MT#979|@MICHAELTRAZZI

DK#200|@DAVIDSKRUEGER

EM#176|@EMOLLICK

GM#157|@GARYMARCUS

MB#20|@MILES_BRUNDAGE

110

Anthropic launches Claude Fable, a Mythos-class model that scored 91 on Every's senior-engineer benchmarkThe model costs twice as much as Claude Opus.

4h|999.3k10.2k939

ZE#1471|@ZEPHYR_Z9

AV#1245|@ALTRYNE

B(#850|@BEFFJEZOS

AC#518|@ANDREWCURRAN_

120

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.

1d|171.8k653252

SS#1967|@STEVESI

MB#1802|@MATTHEWBERMAN

RP#1031|@ROHANPAUL_AI

C🎉#848|@CHARLES_IRL

JC#253|@JULIEN_C

C🤗#67|@CLEMENTDELANGUE

130

Apple to unveil a redesigned Siri with a dedicated app at WWDC 2026The release includes iOS 27 and overhauled Apple Intelligence.

1d|622.3k9.1k293

AV#1245|@ALTRYNE

BT#629|@BENTHOMPSON

140

Keller Jordan, Muon optimizer creator, updates Modded-NanoGPT benchmark, showing Muon and Nesterov-boosted Shampoo beat standard AdamSpectral descent was the slowest, requiring 8,225 steps.

11h|98.4k880310

ET#1859|@TORCHCOMPILED

KM#1625|@KONSTMISH

YJ#829|@YOUJIACHENG

YB#527|@YAROSLAVVB

KJ#426|@KELLERJORDAN0

RP#250|@KHOOMEIK

RA#79|@_AROHAN_

SZ#59|@SUCHENZANG

150

Apple overhauls Apple Intelligence with a new hybrid architecture developed with Google that pairs its foundation models with GeminiWorkloads execute both locally on-device and on remote servers.

13h|612.2k12.6k460

ZE#1471|@ZEPHYR_Z9

AV#1245|@ALTRYNE

KR#1055|@KYLIEBYTES

160

Apple unveils Apple Intelligence-powered Clean Up, Extend, and Reframe photo tools for iOS 27The tools use on-device spatial and generative image models

12h|883.9k8.4k1.5k

170

Apple indefinitely delays its overhauled Siri AI on EU iPhones and iPads, citing DMA security and privacy risksEU regulators rejected Apple's proposed 'Trusted System Agent' intermediary.

13h|533.4k4.1k380

CH#1448|@KIMMONISMUS

180

Yuval Noah Harari warns Javier Milei's AI corporate personhood proposal risks creating uncontrollable statesJoel Leibo argues AI can be sanctioned without legal personhood.

22h|650.9k2k690

MT#449|@TEGMARK

T(#421|@TEORTAXESTEX

MS#161|@MUSTAFASULEYMAN

190

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.

1d|187.2k1.3k138

SÓ#1466|@S_OHEIGEARTAIGH

JL#1151|@JEFFLADISH

LA#975|@SCALING01

⿻A#357|@IAMTRASK

N✈️#305|@NILOOFAR_MIRE

200

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.

15h|141.2k569273

ND#1871|@NICKADOBOS

TC#1053|@TUHINCHAKR

LL#500|@LUCY3_LI

MI#366|@MOHITIYYER

EM#176|@EMOLLICK

210

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.

17h|271.4k1.8k200

SS#1967|@STEVESI

ZE#1471|@ZEPHYR_Z9

CH#1448|@KIMMONISMUS

RP#1031|@ROHANPAUL_AI

AC#518|@ANDREWCURRAN_

220

Siheng Zhao's LadderMan enables zero-shot sim-to-real ladder climbing and manipulation on Unitree G1 humanoid robotsVFM depth mapping guides the robot's real-world movements.

15h|72.8k36487

GS#1523|@GUANYASHI

BI#1366|@BILALTWOVEC

CP#737|@CHRIS_J_PAXTON

PC#719|@PETERXICHEN

RS#323|@SCOBLEIZER

EJ#36|@ERICJANG11

230

Xiaomi and TileRT_AI launch MiMo-V2.5-Pro-UltraSpeed, hitting 1,000 tokens per second on a 1-trillion-parameter MoE using eight standard GPUsThe system uses FP4 quantization and DFlash speculative decoding.

16h|322.9k2.3k886

ZE#1471|@ZEPHYR_Z9

GB#1275|@GAVINSBAKER

YJ#829|@YOUJIACHENG

JH#38|@JEREMYPHOWARD

240

OpenAI updates ChatGPT with interactive data visualization charts, full-screen writing mode, and paid model effort controlsLong chat histories now receive automatic tables of contents.

15h|325.1k4k774

TI#758|@THSOTTIAUX

250

OpenAI releases AGI roadmap targeting March 2028 for AI systems to perform a significant fraction of researchCritics call the plan PR after Altman's lobbying reports.

9h|136k1.2k246

ZE#1471|@ZEPHYR_Z9

CH#1448|@KIMMONISMUS

WH#1409|@NREHIEW_

RP#1031|@ROHANPAUL_AI

LA#975|@SCALING01

260

Apple rolls out Apple Intelligence upgrades enabling Siri to parse onscreen content, understand personal context, and control appsThe update integrates web-browsing tools for broader world knowledge

13h|652.1k9.9k380

CH#1448|@KIMMONISMUS

270

Apple announces Siri voice customization at WWDC26 with sliders to adjust assistant pace and expressivityThe settings apply system-wide across Maps and Safari.

13h|454.7k9.8k194

KK#1181|@KEVINAKWOK

280

Apple's most powerful on-device AI model requires at least 12GB of memory and M3, M4, or iPhone 17 Pro hardwareThe features will not initially launch in the EU.

12h|1.9M3.8k529

CH#1448|@KIMMONISMUS

290

Google cuts its AI Plus subscription price to $4.99 and doubles cloud storage to 400GBThe plan previously cost $7.99 with 200GB of storage.

13h|788.5k2.5k383

👩‍💻P#270|@DYNAMICWEBPAIGE

300

Unsloth creator Daniel Han and Ben Burtenshaw transition OpenEnv to community ownership to standardize agentic reinforcement learningThe protocol decouples AI training harnesses from RL environments.

16h|96.5k451236

C🎉#848|@CHARLES_IRL

DH#773|@DANIELHANCHEN

WB#340|@WILLCCBB

C🤗#67|@CLEMENTDELANGUE

TW#17|@THOM_WOLF