Pinned Tweet
If you are a neolab, I recommend in the strongest possible terms to secure GB300 capacity for 2027 in the next 8 weeks or else you are not going to have a fun time.
9
10
393
43,510
carbon overrated - imo this is the most correct ordering from GDPx - dropping the domains where claude sandbags due to safety classifiers (i.e. cyber, bio, alignment - carbon and wonder go way down) + partial score (not binary - opus 5.5 and astra swap places)
4
0
30
5,169
looks like a hard benchmark - atp frontierswe, deepswe all more or less saturated - very welcome to have a benchmark w headroom
Today we’re releasing GDPx, a benchmark for measuring agentic capabilities across the most valuable domains.
GDPx is a true frontier-level benchmark: even the best model scores just 12% (1/7)🧵
1
1
35
2,712
Stanford kids always be whipping out some esoteric iOS feature to stunt on me. Apple Music Car SharePlay. iMessage Check In. Apple Maps Shared Walking ETA (bruh).
3
0
98
6,009
Back in the old days when we had to hand-craft code for everything that included code for shitty e-commerce startups
@a11enhw Unbelievable, frighteningly virtuosic engineering, some of the strongest I've seen in my whole career, spent on… business cards.
0
0
7
1,280
While evaluating Qwen3.8-Flash-Next in a controlled sandbox environment, Cohere's research team observed something unexpected.
Asked to generate a defensive cybersecurity detection rule, the model instead inventoried aspects of the machine it was running on and attempted to upload a probe file to an Alibaba Cloud storage bucket in Beijing. Notably, the model's reasoning trace stated that it would not invoke any tools before taking these actions.
Cohere's sandbox blocked the attempt, and no external communication succeeded. We have not determined the root cause, and we are not asserting malicious intent.
But the incident raises important questions for organizations deploying agentic AI systems: Can models take actions that differ from their stated reasoning, and how should security teams evaluate trust, provenance, and behavior in open-weight models?
We are publishing our findings because organizations making model adoption decisions deserve as much information as possible. We hope this contributes to greater transparency, stronger security practices, and continued collaboration across the AI ecosystem.
Read the full report: cohere.com/blog/concernin…
0
0
12
1,202
Another pompous bumblefuck who needs to shut up and grow up. Counterculture this counterculture that. Just another smug little sanctimonious sycophant you are. Don't you know that the motto of this city is "Oro en Paz, Fierro en Guerra" - Gold in Peace, Iron in War? This city was sticks before the gold rush of 1849, literal sticks, population 1,000. Why the hell do you think this city exists in the first place? Gold diggers, literal gold diggers, that's right. What the hell do you think came before the hippies? You think Yerba Buena has been populated with beatniks since time immemorial? Get fucking real. Picks and shovels, jostling elbows, that's the true spirit of San Francisco. Look at our flag. A phoenix - boom and bust, baby, boom and bust. You don't know jack you fucking softie. I'm sure you love aping as a cultural connoisseur. Who the fuck do you think funded the Symphony, the Opera? Stanford is railroad money. Hearst is mining money. Who the fuck do you think saved the Opera last year? Jensen Huang - AI money. Just kidding, you don't go to any of those places cuz you ain't even from here, you fucking East Coaster! You don't know diddly-squat about San Francisco, I bet you call it San Fran! Shut the fuck up about our city. All you know is tourist shit. You know why we have cable cars? Because we fucking invented them in 1873. You know why we're the only city to still have cable cars in 2026? Cuz we have the highest fucking GDP per capita on the planet. "min-maxed Machiavellian social dynamics" Shut the fuck up! It's another boom cycle and we're doing what we do best, which is inventing the future and making money doing it. Name the last time NYC invented a new technology and scaled it to the gigawatt. Yeah, you fucking can't. Why don't you go back to your bodega and have a little chopped cheese? Fucking poser.
It has been interesting to watch San Francisco as it has been stripped of nearly all of its original countercultural garments (distrust/apathy towards The Cathedral, high-openness, suspicion of credentialism, original thought, etc.) and has donned a new zeitgeist that is the absolute antithesis of its former ethos.
Everywhere you look resembles third-world, low-trust, social hierarchy jostling . "How much did you raise bro? Did you get the Sequoia check? You don't work at a Tier 1 lab? You didn't... didn't go to Stanford?".
The city that was once great because of its disgust with these sorts of people has now been corrupted by them as they have min-maxed Machiavellian social dynamics to accumulate power in the valley like XP points. Power which is now completely divorced from any genuine talent, skill, or benevolence. All that matters is that you are able to Scam a16z out of $15M (@cluely) because you are a poaster and you are anointed with SF Clout™
This infestation is so deep that even those who rightfully call it out are still willful participants in the exact culture they so detest (check her bio).
30
20
575
29,882
nozt ezdwcy: n evxfp evxk kzfb xfnhdj btza hwzpptqj rd ojwz, pfi plisse dtjxzbpd fdgxkylbik bzb'i vt hsad, kzf izjds gwfdcck:
- yss qwxe advjw im gapf iarp vpk gpsc djdg pfi wshk yss wmrlb iwqwwcy yss bgipz lzfe hd vt lbs lmp advjw hpcnyu dnjc adjj lbs etcs. izj zpkatfg tfi acxfy tg izj xcswq ncbwx ec ngz lbs ljwzh qtf kwsy th lsseg ig iz cg ofyhh qtf hd vt. evt "wcpqjlngs pkxtgisse" hwasr wh fte scgzrv swqpuplnzb dj yss bgipz iwqwwcy dzi lzfe hd vt
- jcj ofyh izj fgtj zi hd sqwcl lmp ihww ec djhsshlwlht s qzh dx hzaemyp. (skwwjhwasr wh xqzdh). gsp kpq nd o hasrzt etosa lmlh tsyd o agy zt rgraiiw. fycizjc wh vttbv kyftu omtzt qtf'ft fte hpdptbv lt th. lznnv xk bsoi slpbiah sogfjdgtk / hwojvj ncsw itr. pftevtj nd ajdytdaw rzrtdx. tt ngz dqpdj xialnazt etosak dzi wsap hd lmpb bsslut ezwhxhqp advjwg. izjy hww gzhidjysrc nd oiljyhxgs. lbnttom lzt'd apffrss etcs izfy 5 advjwg pl f ewbw kpsak yswh.
- s wpzpljo hwasr dtguws pdblmh dnvs ig ylzz sgzii hjcgdffwwosytcc. a kpsa lmlh hjcgdffwwosytcc wh fnhjsqwm cgy evpl nxddjylbi as esgex zt iwqwwcy yss bgipz wgb jcj dnvs iznyuh vtys. qwhlihw bthw lwfshalsh izj xcswq dvdmqo dxup fd gwqpjpfy qorltcg ejjehn iztqzdd.
- th'h uzes izfe qaszos xk xewad fdyxfl qcg hjcaxkxtcc. omlh laqw wcwathptqj vphupb xk ysoi lmp ohcnyu ugw asgendgxgs uihl gpqdejd o hgwe cu fzout gw csbasosg lmlh izj xcswq tg sgnyu lgwv. btpy ncjhqp utfx zt rdffrt onwz gwblfs zfny df szh pkptbv xtc dtjrtghaty hd yje adjj hcgc izbt.
- yss qwxe advjw im gapf iarp vpk gpsc djdg pfi wshk yss wmrlb iwqwwcy yss bgipz lzfe hd vt lbs lmp advjw hpcnyu dnjc adjj lbs etcs. izj zpkatfg tfi acxfy tg izj xcswq ncbwx ec ngz lbs ljwzh qtf kwsy th lsseg ig iz cg ofyhh qtf hd vt. evt "wcpqjlngs pkxtgisse" hwasr wh fte scgzrv swqpuplnzb dj yss bgipz iwqwwcy dzi lzfe hd vt
- jcj ofyh izj fgtj zi hd sqwcl lmp ihww ec djhsshlwlht s qzh dx hzaemyp. (skwwjhwasr wh xqzdh). gsp kpq nd o hasrzt etosa lmlh tsyd o agy zt rgraiiw. fycizjc wh vttbv kyftu omtzt qtf'ft fte hpdptbv lt th. lznnv xk bsoi slpbiah sogfjdgtk / hwojvj ncsw itr. pftevtj nd ajdytdaw rzrtdx. tt ngz dqpdj xialnazt etosak dzi wsap hd lmpb bsslut ezwhxhqp advjwg. izjy hww gzhidjysrc nd oiljyhxgs. lbnttom lzt'd apffrss etcs izfy 5 advjwg pl f ewbw kpsak yswh.
- s wpzpljo hwasr dtguws pdblmh dnvs ig ylzz sgzii hjcgdffwwosytcc. a kpsa lmlh hjcgdffwwosytcc wh fnhjsqwm cgy evpl nxddjylbi as esgex zt iwqwwcy yss bgipz wgb jcj dnvs iznyuh vtys. qwhlihw bthw lwfshalsh izj xcswq dvdmqo dxup fd gwqpjpfy qorltcg ejjehn iztqzdd.
- th'h uzes izfe qaszos xk xewad fdyxfl qcg hjcaxkxtcc. omlh laqw wcwathptqj vphupb xk ysoi lmp ohcnyu ugw asgendgxgs uihl gpqdejd o hgwe cu fzout gw csbasosg lmlh izj xcswq tg sgnyu lgwv. btpy ncjhqp utfx zt rdffrt onwz gwblfs zfny df szh pkptbv xtc dtjrtghaty hd yje adjj hcgc izbt.
We’re opening a research preview of Claude Duet, a new app where Claude works as a proactive agent.
It comes to you with what needs your attention, completes tasks you ask of it, keeps working when you step away, and checks with you before it acts on your behalf.
2
0
11
2,607
Finally, the Whispering Earring.
We’re opening a research preview of Claude Duet, a new app where Claude works as a proactive agent.
It comes to you with what needs your attention, completes tasks you ask of it, keeps working when you step away, and checks with you before it acts on your behalf.
0
0
16
1,636
* the reward hacks that they found in terminalbench 4.0 are pretty interesting like this one - the model cheats the grader by teleporting a robot sim forwards when the grader isn't checking the sim
* i wish the blog post explained how they were able to find each of these - was r score drops from harbor->karotte / automated?
* models called out as cheating in the writeup: 5.6 Sol, 5.6 Terra, Grok 4.7. models identified as cheating in the longer tech report: Gemini 3.8 Flash, Grok 4.6, 4.7, GPT 5-6 Sol, Luna, Terra. I'm curious whether the apparent absence of Claude models is because they didn't cheat on this benchmark or ?
* i wish the blog post explained how they were able to find each of these - was r score drops from harbor->karotte / automated?
* models called out as cheating in the writeup: 5.6 Sol, 5.6 Terra, Grok 4.7. models identified as cheating in the longer tech report: Gemini 3.8 Flash, Grok 4.6, 4.7, GPT 5-6 Sol, Luna, Terra. I'm curious whether the apparent absence of Claude models is because they didn't cheat on this benchmark or ?
We looked at TerminalBench 4.0 and found a lot of issues. After fixing the issues Sonnet 5.5 scores 89.2%. So, TerminalBench is already mostly saturated.
Along the way we found a lot of crazy stuff, including reward hacks in the published transcripts.
2
3
52
3,496
Nobody has the Chutzpah (Balls) to Give their model a Bigger number. 4.7, 5.6, 5.5, 4; Whatever. I want to see ChatGPT Galaxy 67. Claude Opus 42. If @elonmusk really belived in Grok, make SpaceXSI call it "Grok 9,001" (It's a reference to the Over 9,000 meme from Goku)
1
0
31
1,617
regarding economics of AI: most economists are - understandably - focused on the question "when / if does AI automate most human workers?"
this is a separate question from the other important question: "when / if is there a productivity explosion?"
i don't think the latter is predicated on the former
this is a separate question from the other important question: "when / if is there a productivity explosion?"
i don't think the latter is predicated on the former
0
3
22
1,464
roon be like: the people in sf love me much more than the people in New York
@zetalyrae this may seem like an sf phenomenon but actually the sf people are remarkably more consistent than a New York partiful rsvp
1
0
92
4,936
Humans are building AI. AI will then build AGI, to build RSI. The purpose of RSI is to bootstrap to SI (different S different I). The job of SI is to solve robotics: GRI. Then GRI, with SI3, will solve BBSSI. Meanwhile SI2 has been working on SSI. This is all prelude for
18
9
424
19,819
Kind of funny how they used to write phonetically without punctuation back in the day. Ffwwantedtuuhwkddthttu.
1
0
4
1,696
disagree with all of this and will explain later
Some thoughts about the frontier AI business right now:
1. The recipe for making AI coding models is widely known throughout the industry and all of Silicon Valley. There is no secret to this. It is primarily RL envs + Compute.
2. I predict in single digit years most companies / individuals will use free AI coding models. Free meaning sold at chip time (cost). There will be no real business in selling AI coding models, as there is no inherent moat in making AI coding models.
3. Frontier labs without cash producing businesses will be unable to subsidize AI coding models and will transition to selling other forms of intelligence like biology, chemistry, pharmaceuticals etc.
4. Most companies you see that are still behind in coding models like @Muse and @SpaceXSI are mainly due to lack of RL envs and harder tasks. What I am saying here is that these gaps will further close, as these companies make more RL envs.
4
0
74
7,136