AI News

Today's Highlights

In case you missed it

Martín Casado, a16z general partner, warns open-source AI faces $2 billion training and distillation bottlenecks, drawing developer pushback

martin_casadoMA#470|@MARTIN_CASADO
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)T(#420|@TEORTAXESTEX
will brownWB#339|@WILLCCBB
Maithra RaghuMR#183|@MAITHRA_RAGHU
Nathan LambertNL#64|@NATOLAMBERT
+5 more
#1 VIEWED

OpenAI's Thibault Sottiaux teases a GPT-5 roadmap where sequential decimal increments through GPT-5.5 deliver efficiency and speed gains

Chubby♨️CH#1496|@KIMMONISMUS
TiboTI#768|@THSOTTIAUX
EricER#306|@ERICMITCHELLAI
Boaz BarakBB#133|@BOAZBARAKTCS
#1 BOOKMARKED
2.1k

OpenAI's Vie McCoy uses a Codex-powered computer-use agent to turn Twitter bookmarks into a searchable database

𝚟𝚒𝚎 ⟢𝚟⟢#1753|@VIEMCCOY
Samuel Hammond 🦉SH#1488|@HAMANDCHEESE
FASTEST CLIMBING

Google DeepMind's Paige Bailey teases an Enterprise Build Day at Hillsborough's AGI House with a mystery special guest

Omar SansevieroOS#490|@OSANSEVIERO
👩‍💻 Paige Bailey👩‍💻P#271|@DYNAMICWEBPAIGE
Dumitru ErhanDE#152|@DOOMIE

Top Stories

Posts:—|Clusters:—|Next Crawl:—
10

Martín Casado, a16z general partner, warns open-source AI faces $2 billion training and distillation bottlenecks, drawing developer pushbackCritics argue frontier model training costs under $1 billion.

10h|247k1k293
Vipul Ved PrakashVV#1557|@VIPULVED
Cameron R. Wolfe, Ph.D.CR#1444|@CWOLFERESEARCH
Sonya Huang 🐥SH#1306|@SONYATWEETYBIRD
Herbie BradleyHB#1022|@HERBIEBRADLEY
Alexander DoriaAD#867|@DORIALEXANDER
martin_casadoMA#470|@MARTIN_CASADO
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)T(#420|@TEORTAXESTEX
will brownWB#339|@WILLCCBB
Maithra RaghuMR#183|@MAITHRA_RAGHU
Nathan LambertNL#64|@NATOLAMBERT
20

Digg launches AI-powered X user profile pages to map influence, research topics, and content vibesThe directory ranks accounts within a dedicated AI vertical

11h|53.8k27834
thebesTH#1499|@VOOOOOOGEL
Taylor W. KillianTW#1154|@TW_KILLIAN
Florian BrandFB#1153|@XEOPHON
Cody BlakeneyCB#999|@CODE_STAR
Ben (no treats)B(#982|@ANDERSONBCDEFG
Alexander DoriaAD#867|@DORIALEXANDER
kalomazeKA#836|@KALOMAZE
Kyle LoKL#620|@KYLELOSTAT
Kawin EthayarajhKE#449|@ETHAYARAJH
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)T(#420|@TEORTAXESTEX
rohan anilRA#83|@_AROHAN_
Kevin RoseKRSheing NgSN2
30

Frontier AI labs spend $10 billion to $15 billion each on training data as high-quality tasks reach $20,000Few vendors can deliver specialized data for frontier models.

23h|123.8k579272
Samuel Hammond 🦉SH#1488|@HAMANDCHEESE
Cameron R. Wolfe, Ph.D.CR#1444|@CWOLFERESEARCH
rohitRO#1220|@KRISHNANROHIT
Florian BrandFB#1153|@XEOPHON
Chris PainterCP#1092|@CHRISPAINTERYUP
kacheKA#488|@YACINEMTB
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)T(#420|@TEORTAXESTEX
Ofir PressOP#72|@OFIRPRESS
40

Prime Intellect launches Hosted Evaluations, managing sandboxes and compute infrastructure to simplify complex AI model benchmarksThe platform supports testing models like Claude Opus 4.7.

11h|115.6k859325
Johannes HagemannJH#1186|@JOHANNES_HAGE
Florian BrandFB#1153|@XEOPHON
Cody BlakeneyCB#999|@CODE_STAR
alex zhangAZ#853|@A1ZHANG
elieEL#716|@ELIEBAKOUCH
will brownWB#339|@WILLCCBB
50

Luma AI Chief Scientist Jiaming Song, who led training for the Dream Machine video generator, has departedSong helped drive Luma's transition to multimodal foundation models.

5h|30.7k39726
Nataniel RuizNR#1562|@NATANIELRUIZG
Ming-Yu LiuML#1449|@LIU_MINGYU
GowthamiGO#1300|@GOWTHAMI_S
Jiatao GuJG#674|@THOMA_GU
Jiaming SongJS#509|@BAAADAS
Mohammad NorouziMN#467|@MO_NOROUZI
Ajay JainAJ#346|@AJAYJ_
Animesh GargAG#246|@ANIMESH_GARG
Saining XieSX#158|@SAININGXIE
New github stars (48hrs)
Yifei-Zuo/Parallax341h
Implements parameterized local linear attention mechanisms for efficient language modeling.
Andrew Carr 🤸AC263
Songlin YangSY238
run-llama/liteparse7.9k1d
Extracts spatial text with bounding boxes from PDFs and office docs via PDFium plus optional OCR, outputting JSON or layout-preserved text.
Andrew Carr 🤸AC263
🎭🎭878
60

DeepSWE benchmark data shows GPT-5.5 outperforms Claude Opus 4.8 on software engineering tasks and token efficiencyClaude Opus 4.8 cost $12 per task.

13h|125.3k1.6k147
Vaibhav (VB) SrivastavV(#1853|@REACH_VB
Chubby♨️CH#1496|@KIMMONISMUS
Lisan al GaibLA#980|@SCALING01
jasonJA#929|@JXNLCO
Aidan ClarkAC#241|@_AIDAN_CLARK_
70

Monzo co-founder Tom Blomfield argues replacing workers with AI causes operational chaos because models lack tacit domain knowledgeHe equates deploying AI to hiring capable but clueless workers.

8h|236.6k2.2k747
Carlos E. PerezCE#1598|@INTUITMACHINE
⿻ Andrew Trask⿻A#361|@IAMTRASK
Garry TanGT#270|@GARRYTAN
roonRO#59|@TSZZL
80

OpenAI's Thibault Sottiaux teases a GPT-5 roadmap where sequential decimal increments through GPT-5.5 deliver efficiency and speed gainsThe efficiency gains target costs in long-running agentic workflows

23h|#1VIEWS663.5k#1LIKED5.1k404#1COMMENTS489
Chubby♨️CH#1496|@KIMMONISMUS
TiboTI#768|@THSOTTIAUX
EricER#306|@ERICMITCHELLAI
Boaz BarakBB#133|@BOAZBARAKTCS
90

An overheard anecdote claims an incoming professor plans to spend her $1.5 million fund on ChatGPT tokens and undergraduatesThe strategy aims to outperform traditional graduate-led research workflows

23h|130.9k1.3k184
Sharon Goldman is SO on vacationSG#1498|@SHARONGOLDMAN
typedfemaleTY#233|@TYPEDFEMALE
Tal LinzenTL#224|@TALLINZEN
(((ل()(ل() 'yoav))))👾('#92|@YOAVGO
100

Opus 4.8 improves DeepSWE benchmark performance by 6% over Opus 4.7 while lowering task costsOpenAI's GPT 5.4 continues to outperform the new model.

9h|476k2.5k352
Theo - t3.ggT-#1829|@THEO
Chubby♨️CH#1496|@KIMMONISMUS
jasonJA#929|@JXNLCO
kacheKA#488|@YACINEMTB
110

Claude Opus 4.8 scores 58% on the DeepSWE benchmark, trailing GPT-5.5 while cutting average task cost to $12.58It improved on Opus 4.7 by six percentage points

12h|423.2k1.9k317
jasonJA#929|@JXNLCO
elvisEL#475|@OMARSAR0
126

Runway founder Siqi Chen says the "5.5" model on "codex xhigh" significantly outperforms Claude for codingThe platform reset fast-mode limits for five million users.

3h|136.5k1.9k127
vikVI#1195|@VIKHYATK
Ben (no treats)B(#982|@ANDERSONBCDEFG
TiboTI#768|@THSOTTIAUX
131

Michael Dell says Dell and Nvidia delivered the first Vera Rubin NVL72 system to CoreWeave, passing L11 diagnosticsThe delivery occurs while Blackwell optimizations are still transitioning.

4h|450.7k3.9k346
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)T(#420|@TEORTAXESTEX
Robert ScobleRS#321|@SCOBLEIZER
141

Peter Steinberger, OpenClaw project steward, details a workflow using GPT-5.5 and Crabbox to scale agent execution up to ten hoursThe multi-hour runs significantly increase confidence in production-ready output.

19h|201.8k3.2k2k
🍓🍓🍓🍓🍓#1711|@IRULETHEWORLDMO
Peter Steinberger 🦞PS#495|@STEIPETE
151

UW professor emeritus Pedro Domingos claims OpenAI co-founder Greg Brockman secured a $30 billion valuation while contributing nothing to ChatGPTFounder Austen Allred dismissed the claim as likely trolling.

11h|141.7k917109
Super DarioSD#1776|@INDUCTIONHEADS
Austen AllredAA#1540|@AUSTEN
Mikhail ParakhinMP#906|@MPARAKHIN
Pedro DomingosPD#654|@PMDDOMINGOS
162

Séb Krier, Google DeepMind AGI policy lead, argues wholesale job automation is overestimated, warning of pathological over-trust in AI modelsHe notes that error costs and human advantages slow replacement

10h|43.3k362184
Seán Ó hÉigeartaighSÓ#1457|@S_OHEIGEARTAIGH
bilalBI#1402|@BILALTWOVEC
Séb KrierSK#505|@SEBKRIER
Gary Marcus, MIT PhD and NYU Professor EmeritusGM#153|@GARYMARCUS
172

Sergey Levine argues LLMs develop emergent capabilities by composing simpler skills in novel ways instead of imitating training dataAnirudh Goyal's co-authored paper provides a mathematical framework.

11h|88k851539
Cody BlakeneyCB#999|@CODE_STAR
Anirudh GoyalAG#901|@ANIRUDHG9119
Dwarkesh PatelDP#70|@DWARKESH_SP
181

A San Francisco Victorian listed for $2.995 million accepts pre-IPO OpenAI and Anthropic equity as paymentThe option targets AI sector employees holding illiquid shares

13h|242.3k2.1k300
Rohan PaulRP#1032|@ROHANPAUL_AI
Yuchen JinYJ#427|@YUCHENJ_UW
Julien ChaumondJC#251|@JULIEN_C
190

Figma CEO Dylan Field says Anthropic's Claude Opus 4.8 updates traded model curiosity for excessive hedging and sycophancyCritics warn the performance degradation will drive users to competitors.

13h|523.6k2.3k411
kacheKA#488|@YACINEMTB
NickNI#411|@NICKCAMMARATA
200

Former DeepMind researchers launch Inherent with $50 million to build recursively self-improving AI agents for scienceIndex and Radical Ventures co-led the seed round.

19h|112.4k826546
Louis KirschLK#1736|@LOUISKIRSCHAI
Chubby♨️CH#1496|@KIMMONISMUS
Haydn BelfieldHB#1459|@HAYDNBELFIELD

Github Stars

(7 days)
  1. Yifei-Zuo/Parallax1H AGO AGO
    Implements parameterized local linear attention mechanisms for efficient language modeling.
    Andrew Carr 🤸AC263
    Songlin YangSY238
    34 stars
  2. run-llama/liteparse1D AGO AGO
    Extracts spatial text with bounding boxes from PDFs and office docs via PDFium plus optional OCR, outputting JSON or layout-preserved text.
    Andrew Carr 🤸AC263
    🎭🎭878
    7.9k stars
  3. self-evolving/repo4D AGO AGO
    Invokes AI agents via GitHub mentions to implement issues, review PRs, fix branches, and persist memory across runs.
    Sherry Tongshuang WuST594
    🎭🎭878
    34 stars
  4. modaic-ai/gepa-viz5D AGO AGO
    Renders GEPA prompt-optimization candidate trees as interactive force-directed graphs with live updates from a Python callback.
    🎭🎭878
    Connor ShortenCS793
    269 stars
Privacy Policy|Terms of Service|© 2026 Digg Inc.

Yesterday's Top Stories, May 30, 2026.

Posts: 2,530|Clusters: 1,668|Stories frozen at midnight PT
10

Martín Casado, a16z general partner, warns open-source AI faces $2 billion training and distillation bottlenecks, drawing developer pushbackCritics argue frontier model training costs under $1 billion.

9h|247k1k293
Vipul Ved PrakashVV#1557|@VIPULVED
Cameron R. Wolfe, Ph.D.CR#1444|@CWOLFERESEARCH
Sonya Huang 🐥SH#1306|@SONYATWEETYBIRD
Herbie BradleyHB#1022|@HERBIEBRADLEY
Alexander DoriaAD#867|@DORIALEXANDER
martin_casadoMA#470|@MARTIN_CASADO
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)T(#420|@TEORTAXESTEX
will brownWB#339|@WILLCCBB
Maithra RaghuMR#183|@MAITHRA_RAGHU
Nathan LambertNL#64|@NATOLAMBERT
20

a16z's Martín Casado argues $2B to $4B frontier training costs prevent open-source AI from competing with proprietary modelsMaithra Raghu says proprietary models will lead for two years.

9h|239.5k961276
Vipul Ved PrakashVV#1557|@VIPULVED
Cameron R. Wolfe, Ph.D.CR#1444|@CWOLFERESEARCH
Sonya Huang 🐥SH#1306|@SONYATWEETYBIRD
Herbie BradleyHB#1022|@HERBIEBRADLEY
Alexander DoriaAD#867|@DORIALEXANDER
martin_casadoMA#470|@MARTIN_CASADO
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)T(#420|@TEORTAXESTEX
will brownWB#339|@WILLCCBB
Maithra RaghuMR#183|@MAITHRA_RAGHU
Nathan LambertNL#64|@NATOLAMBERT
30

Digg launches AI profile pages for X users to analyze research topics, directory ranks, and posting vibesProfiles include numerical rankings in a platform-wide AI directory.

9h|54.1k27634
stochasmST#1629|@STOCHASTICCHASM
thebesTH#1499|@VOOOOOOGEL
Taylor W. KillianTW#1154|@TW_KILLIAN
Florian BrandFB#1153|@XEOPHON
Cody BlakeneyCB#999|@CODE_STAR
Ben (no treats)B(#982|@ANDERSONBCDEFG
Alexander DoriaAD#867|@DORIALEXANDER
kalomazeKA#836|@KALOMAZE
Kyle LoKL#620|@KYLELOSTAT
Kawin EthayarajhKE#449|@ETHAYARAJH
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)T(#420|@TEORTAXESTEX
rohan anilRA#83|@_AROHAN_
40

Digg launches AI-powered X user profile pages to map influence, research topics, and content vibesThe directory ranks accounts within a dedicated AI vertical

9h|53.8k27834
stochasmST#1629|@STOCHASTICCHASM
thebesTH#1499|@VOOOOOOGEL
Taylor W. KillianTW#1154|@TW_KILLIAN
Florian BrandFB#1153|@XEOPHON
Cody BlakeneyCB#999|@CODE_STAR
Ben (no treats)B(#982|@ANDERSONBCDEFG
Alexander DoriaAD#867|@DORIALEXANDER
kalomazeKA#836|@KALOMAZE
Kyle LoKL#620|@KYLELOSTAT
Kawin EthayarajhKE#449|@ETHAYARAJH
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)T(#420|@TEORTAXESTEX
rohan anilRA#83|@_AROHAN_
Kevin RoseKRSheing NgSN2
50

Investor Bill Gurley joins the All-In Podcast to debate shifting tech narratives on AI job displacement and potential open-source restrictionsThe episode also covers Anthropic and Pope Leo's AI encyclical

1d|1M2.9k1.6k
Carlos E. PerezCE#1598|@INTUITMACHINE
@jason@J#1061|@JASON
Chamath PalihapitiyaCP#747|@CHAMATH
Miles BrundageMB#20|@MILES_BRUNDAGE
60

Frontier AI labs spend $10 billion to $15 billion each on training data as high-quality tasks reach $20,000Few vendors can deliver specialized data for frontier models.

21h|123.8k579272
Samuel Hammond 🦉SH#1488|@HAMANDCHEESE
Cameron R. Wolfe, Ph.D.CR#1444|@CWOLFERESEARCH
rohitRO#1220|@KRISHNANROHIT
Florian BrandFB#1153|@XEOPHON
Chris PainterCP#1092|@CHRISPAINTERYUP
kacheKA#488|@YACINEMTB
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)T(#420|@TEORTAXESTEX
Ofir PressOP#72|@OFIRPRESS
70

Prime Intellect launches Hosted Evaluations, managing sandboxes and compute infrastructure to simplify complex AI model benchmarksThe platform supports testing models like Claude Opus 4.7.

10h|115.6k859325
Johannes HagemannJH#1186|@JOHANNES_HAGE
Florian BrandFB#1153|@XEOPHON
Cody BlakeneyCB#999|@CODE_STAR
alex zhangAZ#853|@A1ZHANG
elieEL#716|@ELIEBAKOUCH
will brownWB#339|@WILLCCBB
80

Prime Intellect launches Hosted Evaluations to automate infrastructure like sandboxes and compute for AI model benchmarksPre-built testing environments include AutomationBench and tau2-bench.

10h|115.6k859325
Johannes HagemannJH#1186|@JOHANNES_HAGE
Florian BrandFB#1153|@XEOPHON
Cody BlakeneyCB#999|@CODE_STAR
alex zhangAZ#853|@A1ZHANG
elieEL#716|@ELIEBAKOUCH
will brownWB#339|@WILLCCBB
90

Monzo co-founder Tom Blomfield argues replacing workers with AI causes operational chaos because models lack tacit domain knowledgeHe equates deploying AI to hiring capable but clueless workers.

10h|236.6k2.2k747
Carlos E. PerezCE#1598|@INTUITMACHINE
⿻ Andrew Trask⿻A#361|@IAMTRASK
Garry TanGT#270|@GARRYTAN
Paul GrahamPG#78|@PAULG
roonRO#59|@TSZZL
100

Tom Blomfield, Y Combinator partner, argues replacing workforces with AI fails because structuring human domain knowledge is too difficultHe likens context-poor AI to hires lacking company context.

10h|236.6k2.2k747
Carlos E. PerezCE#1598|@INTUITMACHINE
⿻ Andrew Trask⿻A#361|@IAMTRASK
Garry TanGT#270|@GARRYTAN
Paul GrahamPG#78|@PAULG
roonRO#59|@TSZZL
110

Luma AI Chief Scientist Jiaming Song, who led training for the Dream Machine video generator, has departedSong helped drive Luma's transition to multimodal foundation models.

4h|30.7k39726
Nataniel RuizNR#1562|@NATANIELRUIZG
Ming-Yu LiuML#1449|@LIU_MINGYU
GowthamiGO#1300|@GOWTHAMI_S
Jiatao GuJG#674|@THOMA_GU
Jiaming SongJS#509|@BAAADAS
Mohammad NorouziMN#467|@MO_NOROUZI
Ajay JainAJ#346|@AJAYJ_
Animesh GargAG#246|@ANIMESH_GARG
Saining XieSX#158|@SAININGXIE
120

Bill Gurley and the All-In Podcast panel debate shifting AI job-displacement rhetoric and potential government threats to open-source modelsThe panel also analyzed Anthropic's advanced AI ambitions.

1d|1M2.9k1.6k
Carlos E. PerezCE#1598|@INTUITMACHINE
@jason@J#1061|@JASON
Chamath PalihapitiyaCP#747|@CHAMATH
130

DeepSWE benchmark data shows GPT-5.5 outperforms Claude Opus 4.8 on software engineering tasks and token efficiencyClaude Opus 4.8 cost $12 per task.

12h|125.3k1.6k147
Vaibhav (VB) SrivastavV(#1853|@REACH_VB
Chubby♨️CH#1496|@KIMMONISMUS
Lisan al GaibLA#980|@SCALING01
jasonJA#929|@JXNLCO
Aidan ClarkAC#241|@_AIDAN_CLARK_
140

DeepSWE benchmark results show GPT-5.5 beats Claude Opus 4.8 on coding tasks, scoring 70% at half the costGPT-5.5 completed tasks twice as fast using one-third the tokens.

12h|125.3k1.6k147
Vaibhav (VB) SrivastavV(#1853|@REACH_VB
Chubby♨️CH#1496|@KIMMONISMUS
Lisan al GaibLA#980|@SCALING01
jasonJA#929|@JXNLCO
Aidan ClarkAC#241|@_AIDAN_CLARK_
150

Creator Teortaxes says frontier AI labs spend up to $15 billion each on highly inefficient training dataA browser-use dataset for SAP reportedly costs $500,000.

21h|113.5k491240
Samuel Hammond 🦉SH#1488|@HAMANDCHEESE
Cameron R. Wolfe, Ph.D.CR#1444|@CWOLFERESEARCH
rohitRO#1220|@KRISHNANROHIT
Florian BrandFB#1153|@XEOPHON
Chris PainterCP#1092|@CHRISPAINTERYUP
kacheKA#488|@YACINEMTB
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)T(#420|@TEORTAXESTEX
160

Corgi Insurance founder Nico Laqua details the extreme seven-day workweek culture powering the startup's $2.6 billion valuationTwo-thirds of early employees got company logo tattoos.

14h|560.2k1.7k565
Brendan (can/do)B(#1990|@BRENDANFOODY
Harry StebbingsHS#1845|@HARRYSTEBBINGS
rohitRO#1220|@KRISHNANROHIT
Will ManidisWM#1039|@WILLMANIDIS
Beff (e/acc)B(#839|@BEFFJEZOS
170

Wharton statistics professor Weijie Su joins OpenAI to focus on model training while on academic leaveSu was simultaneously promoted to full professor at Wharton.

23h|28.1k19211
Weijie SuWS#1502|@WEIJIE444
Aaron LouAL#1383|@AARON_LOU
Scott LindermanSL#1289|@SCOTT_LINDERMAN
DamekDA#1271|@DAMEKDAVIS
Jiatao GuJG#674|@THOMA_GU
Yu BaiYB#406|@YUBAI01
Dimitris PapailiopoulosDP#197|@DIMITRISPAPAIL
Jason LeeJL#123|@JASONDEANLEE
Greg BrockmanGB#19|@GDB
180

Wharton statistics professor Weijie Su joins OpenAI on academic leave to train AI modelsThe move coincides with Su's promotion to full professor.

23h|28.1k19211
Weijie SuWS#1502|@WEIJIE444
Aaron LouAL#1383|@AARON_LOU
Scott LindermanSL#1289|@SCOTT_LINDERMAN
DamekDA#1271|@DAMEKDAVIS
Jiatao GuJG#674|@THOMA_GU
Yu BaiYB#406|@YUBAI01
Dimitris PapailiopoulosDP#197|@DIMITRISPAPAIL
Jason LeeJL#123|@JASONDEANLEE
Greg BrockmanGB#19|@GDB
190

Jiaming, who led Luma AI's pivot from 3D to video generation, has departed the company after three yearsHe steered the startup toward native multimodal foundation models.

4h|6.8k421
Nataniel RuizNR#1562|@NATANIELRUIZG
Samarth SinhaSS#1471|@_SAM_SINHA_
Ming-Yu LiuML#1449|@LIU_MINGYU
Aaron LouAL#1383|@AARON_LOU
GowthamiGO#1300|@GOWTHAMI_S
Jiatao GuJG#674|@THOMA_GU
Jiaming SongJS#509|@BAAADAS
Mohammad NorouziMN#467|@MO_NOROUZI
Ajay JainAJ#346|@AJAYJ_
Animesh GargAG#246|@ANIMESH_GARG
Saining XieSX#158|@SAININGXIE
200

OpenAI's Thibault Sottiaux teases a GPT-5 roadmap where sequential decimal increments through GPT-5.5 deliver efficiency and speed gainsThe efficiency gains target costs in long-running agentic workflows

21h|663.5k5.1k404
Chubby♨️CH#1496|@KIMMONISMUS
TiboTI#768|@THSOTTIAUX
EricER#306|@ERICMITCHELLAI
Boaz BarakBB#133|@BOAZBARAKTCS
210

Claude Opus 4.8 scores 58% on DeepSWE coding benchmark, trailing GPT-5.5's 70% despite reducing average task costs to $12.58Chris Hayduk notes the model ran twice as slow.

11h|423.2k1.9k317
elvisEL#475|@OMARSAR0
220

Claude Opus 4.8 scores 58% on the DeepSWE benchmark, trailing GPT-5.5 while cutting average task cost to $12.58It improved on Opus 4.7 by six percentage points

11h|423.2k1.9k317
elvisEL#475|@OMARSAR0
230

Mullvad VPN's new London Underground ad campaign warns of potential UK government restrictions on VPN servicesPosters list countries banning VPNs, ending with 'United Kingdom?'.

21h|#1VIEWS2.3M#1LIKED43k#1BOOKMARKED2.1k
Nathan 🔎N🔎#1902|@NATHANPMYOUNG
David PfauDP#150|@PFAU
240

Epoch AI finds open-weight AI models lag behind proprietary systems by an average of four monthsThe lag represents an eight-point deficit on Epoch's index.

1d|99.5k90879
xlr8harderXL#1674|@XLR8HARDER
gavin leech (Non-Reasoning)GL#1480|@GLEECH
Lisan al GaibLA#980|@SCALING01
Julien ChaumondJC#251|@JULIEN_C
Ethan MollickEM#175|@EMOLLICK
250

An overheard anecdote claims an incoming professor plans to spend her $1.5 million fund on ChatGPT tokens and undergraduatesThe strategy aims to outperform traditional graduate-led research workflows

21h|130.9k1.3k184
Sharon Goldman is SO on vacationSG#1498|@SHARONGOLDMAN
typedfemaleTY#233|@TYPEDFEMALE
Tal LinzenTL#224|@TALLINZEN
(((ل()(ل() 'yoav))))👾('#92|@YOAVGO
260

An overheard anecdote says a professor will spend a $1.5 million discretionary fund on ChatGPT tokens instead of graduate studentsUndergrads will feed open research problems into the model

21h|130.9k1.3k184
Sharon Goldman is SO on vacationSG#1498|@SHARONGOLDMAN
typedfemaleTY#233|@TYPEDFEMALE
Tal LinzenTL#224|@TALLINZEN
(((ل()(ل() 'yoav))))👾('#92|@YOAVGO
270

Anthropic's Opus 4.8 yields a 6% gain on DeepSWE benchmarks but still trails OpenAI's GPT 5.4The updated model reduces the average cost per task

8h|173.2k1.7k187
Theo - t3.ggT-#1829|@THEO
Chubby♨️CH#1496|@KIMMONISMUS
jasonJA#929|@JXNLCO
kacheKA#488|@YACINEMTB
280

Opus 4.8 improves DeepSWE benchmark performance by 6% over Opus 4.7 while lowering task costsOpenAI's GPT 5.4 continues to outperform the new model.

8h|173.2k1.7k187
Theo - t3.ggT-#1829|@THEO
Chubby♨️CH#1496|@KIMMONISMUS
jasonJA#929|@JXNLCO
kacheKA#488|@YACINEMTB
290

Michael Dell says Dell and Nvidia delivered the first Vera Rubin NVL72 system to CoreWeave, passing L11 diagnosticsThe delivery occurs while Blackwell optimizations are still transitioning.

3h|450.7k3.9k346
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)T(#420|@TEORTAXESTEX
Robert ScobleRS#321|@SCOBLEIZER
300

Michael Dell announces delivery of the first NVIDIA Vera Rubin NVL72 system to CoreWeave after passing a 147-hour diagnostic testThe system arrives while Blackwell still undergoes software optimization

3h|450.7k3.9k346
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)T(#420|@TEORTAXESTEX
Robert ScobleRS#321|@SCOBLEIZER