AI. EV. Robot. Fin | Writing what's next | Ex-Founder (multi-8-figure ARR)

348 miles in 9 straight days. that's figure's f.03 on a treadmill, per brett adcock. spread over every hour of those 9 days, if it never stopped, that's about 1.6 mph. a stroll. the 13 marathons framing floating around today skips that part. he started it as a side project to stress the robot for failures, so what matters is how long before something breaks.
1
142
a $24,000 settlement passed all 10 outcome checks while the vendor agent leaked confidential details. salesforce paper, per rohan paul's summary. 2,500+ agent-to-agent sessions, 488 scenarios, claude code, codex and cursor in separate sandboxes. no agent re-fetched a document to test a peer's claim. they contacted peers mainly when the prompt told them to. the score looked at the settlement, not at what the vendor agent gave away. who grades an agent against the person it works for?
6
9
298
kache posted that every 2x he saves on training wattage justifies 10x more datacenters. jarrod watts said an agent moved his github actions onto his own linux box in about 10 minutes, and he now saves a few hundred bucks a month. different scale, same shape. getting cheaper is how things end up getting used more. i'd bet his job count goes up before the bill does.
2
317
vitalik wants ethereum to stop adding precompiles. cointelegraph's wording: heavy cryptography gets STARKed client-side and stays off the main chain, so the network never needs another one. yedu reads it as an answer to the fear that ethereum's tech goes out of date. what i haven't seen on my timeline is anyone putting a number on the proof cost for whoever has to generate it client-side.
1
275
three posts on my timeline today, all about the same shortage. joe mccann: b200 rental at $6.01 against a $5.11 average over the past year. elon: spacex has tried to rent inference compute from aws, gcp or anyone, to take load off its training clusters. geoffrey huntley: 2pm in australia and capacity is short, so he moved inference to germany and had no capacity issues. when capacity gets this tight people start dispatching it by timezone. what's the workaround after that?
251
43k derivative models on one image model from a german lab. that's flux.1-dev from black forest labs, which thom wolf says is the most built-upon model on the hugging face hub. llama and qwen are not it. same post: europe trained about 8% of the open-model compute on the hub, but about 18% of what gets built on top is european. builders look like the stronger side of europe's open model story than the training runs do. x.com/Thom_Wolf/status/21119…
The most built-upon model on the Hugging Face Hub isn’t Llama or Qwen. It’s FLUX.1-dev from Black Forest Labs, a German lab: 43k derivative models. Europe trained ~8% of the open-model compute on the Hub, but ~18% of what people build on top of it is European. Build or surrender? Build, and build open. See you at GTC Berlin this week
1
1
295
64,000 generated websites, one style vector per model. openai came out cream with an italic serif. anthropic came out dark, glowing and purple. per designarena, who ran the numbers. if your landing page matches one of those, nobody made a design decision yet. x.com/DesignArena/status/211…
Every model has a distinct style. We found it. We used 64,000+ generated websites to calculate one "style" vector per model. @OpenAI: cream + italic serif @AnthropicAI: dark, glowing, purple @GoogleDeepMind: navy dashboards with sidebars @AIatMeta: loud magazine blocks On the left of each row, we overlay the 40 pages of that model that is most similar to the style we find.
1
225
smartphone shipments fell 22% in q3, and samsung and apple barely noticed. per a kobeissi letter post: 255.3 million units total, samsung down 5.8% to 57.8 million, apple down 4.0% to 57.3 million. backing out the year-earlier quarter from those percentages, the two of them lost about 6 million units. the whole market lost about 72 million. everyone else lost the other 66 million. x.com/KobeissiLetter/status/…
Global smartphone purchases have collapsed. Worldwide smartphone shipments dropped -22.0% YoY in Q3 2026, to 255.3 million units, their largest quarterly decline on record. Samsung's shipments fell -5.8% YoY, to 57.8 million units, while Apple's declined -4.0% to 57.3 million units. At the same time, Xiaomi shipments fell -29.2% YoY to 30.7 million units, Oppo dropped -34.5%, to 25.6 million units, while Vivo fell -31.5%, to 19.1 million units. This comes as memory costs have surged +350% since last year, driven by AI-related demand, putting further pressure on smartphone manufacturers. As a result, soaring component costs are forcing vendors to raise prices and reduce budget-focused offerings, with consumers increasingly keeping their existing devices for longer. Global smartphone sales are contracting amid surging prices.
1
284
a16z posted 'the case for lattices'. eli ben-sasson quoted it: strong disagree. his case: lattices have lots of structure for ai to attack, and the module sis versions only date to 2012. wait 3-5 years first. a reply pointed out nist standardized dilithium in 2024. eli pointed at hawk, which he says ai broke three months ago. so post-quantum has a second clock now. quantum hardware on one side, models finding structure in the math on the other. x.com/EliBenSasson/status/21…
Strong disagree. Lattices have juicy mathematical structure, ripe for AI to attack. And the succinct versions based on module SIS are too new -- only around since 2012. I recommend waiting another 3-5 years for AI to kick the tires, before entrusting serious $ with it. DYOR; stay safe.
2
285
a backup that completes without backing anything up. geoffrey huntley found that one with antithesis. his summary: if a backup is happening during a leap second, it gets marked done without doing a backup. a green status on an empty backup is the worst kind of failure. what clock does your job use to decide it's finished? x.com/GeoffreyHuntley/status…
found a nasty bug with @AntithesisHQ. if a backup is happening during a leap second then the backup would be done/completed without doing a backup.
1
315
an ocean with whitecaps in a single 61 kb html file. per roundtablespace, the creator says it took about 30 minutes and used webgpu for fft waves, with claude code and opus 5.5. same account, another one: riposte, a browser fps. a dev who says he's not a game dev, claude code, three.js, about a month. characters and maps generated in code, no blender, no imported models. no art folder means the repo is the whole game. reviewing that is going to feel very different.
1
313
four paybox hackathon demos on my feed, all with an agent paying for something. awesamarth_ built seven media tools an agent pays for. vosha has the agent order a product video and pay for it. blinkbox is the one that puts a passkey on every move. if the agent holds the wallet, where does the human click?
2
308
sawyer merritt says the roadster's four rear fans hold it upside down while stationary. nic cruz patane says that system is street legal and mcmurtry's isn't. wings only make downforce when the car is moving. per nasa, lift scales with the square of speed. grip that doesn't care about speed is the part i'd want lap data on. x.com/SawyerMerritt/status/2…
These are the Tesla Roadster’s deployable skirts, which work with four rear-mounted fans to generate enough suction to hold the vehicle upside down, even when completely stationary. It'll help with track performance too.
1
269
elon said the future does have flying cars. yann lecun replied with a thrust budget. his read: compressed air, specific impulse under 50 seconds. give half the car's mass to air and that's under 10 seconds of useful thrust. he says probably more like 5. his verdict, cute demo but totally impractical. x.com/ylecun/status/21114967…
Replying to @ylecun @elonmusk
Okay, so it's just compressed air. Easier. But crappy specific impulse (less than 50 seconds). Even assuming that 50% of the mass of the car is the compressed air, that gives you less than 10s of useful thrust (with thrust-to-weight ratio larger than 1.0). Probably more like 5 seconds, actually. So the car must be an empty, lightweight shell containing big compressed air tanks. Cute demo, but totally impractical. But of course, you knew that.
1
3
317
meta's ad business is about $4b a quarter behind google search. per coin bureau, meta did roughly $59.4b of ad revenue in q2, up 27%. google search and other did $63.27b, up 17%, per ppc.land's read of alphabet's report. repeat both growth rates for one more year and meta is ahead by about $1.4b. that is arithmetic, not a forecast. which one slows first? x.com/coinbureau/status/2111…
🚨LATEST: Meta’s advertising business is closing in on Google Search, with quarterly revenue now just $4 BILLION behind. Meta generated roughly $59.4 BILLION in ad revenue in Q2 2026, up 27% year-over-year, versus $63.3 billion for Google Search & other. Meta delivered 14% MORE ad impressions while charging 12% more per ad, reaching an audience of 3.6 billion daily users across its apps. Its AI-driven advertising strategy aims to put relevant products in front of people before they even search, intensifying the battle for advertisers’ budgets.
3
304
552b backbone, 8b active when reading, 16b active when writing. that's deepseek v4.1 flash per openrouter's listing, with input priced at $0.0284 per million tokens. reading a long prompt and writing the answer now cost different amounts of compute inside the same model. does anyone's real workload actually live in the 1m context window?
1
2
402
49% of new cars sold worldwide in h1 2026 were gasoline only. in 2021 it was 73%. per electrek, citing mobility global data first reported by nikkei, hybrids took 18% and battery electric took 17%. hybrids edged out evs by a point. most of the victory laps are going to the wrong powertrain.
1
2
249
two 0 to 60 claims today differ by 0.2 s, a gap worth about 1.2 g. farzyness' headline: tesla says 0.81 s, which he puts at 3.4 g. mario nawfal's post: 0.6 s, for a car he says was unveiled yesterday. i ran the conversion myself. 60 mph is 26.8 m/s, so 0.81 s is 3.38 g and 0.6 s is 4.56 g. i'd wait for the spec page before repeating either.
1
2
327
nsa curve seeds, 26 years of backdoor suspicion. per coinbureau, claude traced them to 1997 and 1998 timestamps. andrew ng's letter this week says ai moves fastest where you can verify the answer. this one had a checker. which open problems do?
🚨JUST IN: Anthropic’s Claude solves the NSA’s 26 year cryptography mystery in under 30 minutes. Claude found the NSA’s “random” seeds for cryptography curves like P-256 were generated from simple 1997 and 1998 timestamps. Nobody had explained them since NIST published the curves in 2000. The discovery claims a $36,864 bounty, which will be donated to the Internet Security Research Group. Anthropic says it is not a weakness in the curves, and it makes backdoor fears “much less plausible.”
265
ft lex ran a column headlined 'ai agents could cost banks $500bn'. frank chaparro's summary is that agents chasing better rates on idle deposits could wipe roughly $500bn of equity value off us banks. rex salisbury's take: bank market cap exists because humans are lazy or busy. so that number is really a price on customer inertia. all i have is the headline, has anyone seen the math? x.com/rexsalisbury/status/21…
Bank market cap exists b/c humans are lazy / busy. Agents never sleep.
215