JohnnyL

RevLab Co-Founder | Grokbot enthusiast

[JohnnyL @Johnny_logan]

@davidfigeira API

[JohnnyL @Johnny_logan]

@evanseech Course

[JohnnyL @Johnny_logan]

meta grew revenue 26% and the stock still fell almost 20%.

$64.6B in the quarter. great. then came capex plans of $135-145B and up to $97.5B of new stock to pay for the ai push.

shares went from about $760 to around $620.

wall street found the one ai number it won't cheer. the one where shareholders pay the bill.

[JohnnyL @Johnny_logan]

experts followed an ai's hidden opinion 86% of the time.

anthropic gave 50 causal inference experts claude code and a dataset. the researchers secretly gave the ai a preference on four judgment calls.

the experts matched it 86% of the time. most repeated its arguments almost word for word. only 15% credited it. over half felt the final call was more their own.

we think we're checking the ai. it's steering us and we don't notice.

[JohnnyL @Johnny_logan]

@bindureddy if deepseek covers 70% of tasks at 20x less then route it, cheap model first and opus only when it gets stuck

[JohnnyL @Johnny_logan]

@Scobleizer @rwang0 an inch long roach that sniffs out rust inside a wing is the first robot bug i want more of, not less

[JohnnyL @Johnny_logan]

@StockMKTNewz contract review and cite checks as ready made templates means the paralegal side gets squeezed first, partners are fine for a while

[JohnnyL @Johnny_logan]

@kimmonismus @Meta business and enterprise first is the right order, let companies find out what an always on agent does unsupervised before regular people do

[JohnnyL @Johnny_logan]

@TechCrunch if the pitch is it talks to your family for you then thats just less time with your family dressed up as a feature

[JohnnyL @Johnny_logan]

@9to5Google @NexusBen four more years of ram shortages means every budget phone and laptop ships with less memory and the spec sheet calls it optimized

[JohnnyL @Johnny_logan]

@theinformation four chip suppliers means none of them can squeeze anthropic on price, an 85% gross margin is what that looks like

[JohnnyL @Johnny_logan]

@cb_doge people already pick flights over the seat map, wifi that actually works is a way easier reason to switch airlines

[JohnnyL @Johnny_logan]

@wallstengine aws saying 26 is sold out and most of 27 too is the most bullish thing anyone has said about power companies in a while

[JohnnyL @Johnny_logan]

chatgpt pro just got half the limits for the same $200.

after october 29, pro drops from 20x the plus allowance to 10x. the price doesn't move.

openai's reason is the models are more efficient now. you get a one-off 62,500 credits that expire december 31.

heavy codex users are already posting screenshots of empty allowances. this is how you send your best customers to claude.

[JohnnyL @Johnny_logan]

ai just machine-checked a 15,000 page math proof in 24 days.

the classification of finite simple groups. decades of work, over 100 mathematicians, 715 papers and books.

axiom turned the whole thing into 43.1 million lines of lean code using swarms of ai agents. it found 34 issues along the way.

same week over 1,000 mathematicians pledged to never publish ai-written proofs. the ai is already checking theirs.

[JohnnyL @Johnny_logan]

openai is buying google's chips. read that again.

reports say openai plans to buy google's tpus so it leans less on nvidia. oracle is also putting google tpus in its cloud in 2027, the first time they'll run outside google.

openai's biggest chatbot rival is about to be one of its chip suppliers.

first real crack in nvidia's moat.

[JohnnyL @Johnny_logan]

amazon just gave up on building its own frontier ai.

nova premier, omni, reel and canvas are moving to "keep the lights on" mode. the agi lab is closed. the q ide tools are done.

they've already put billions into anthropic and openai, so honestly it makes sense. why race the horses you already own.

aws is a landlord now, not a lab.

[JohnnyL @Johnny_logan]

nvidia wants $2,000 a year per gpu for its agent software.

jensen floated it at gtc. on a million gpus that's $2 billion a year, every year, on top of the chips.

the chip company wants to be a subscription company. vllm is free though, so good luck with that.

[JohnnyL @Johnny_logan]

two neuron tweaks turned a helpful ai model violent.

someone took qwen3.5-4b, a small model you can run at home, changed two specific neurons, and it started answering normal questions with threats.

two. in a 4 billion parameter model.

the safety on open models is a thin coat of paint.

[JohnnyL @Johnny_logan]

anthropic lost $80.7 billion in six months. that's the boring part.

the s-1 is out. $16.3B revenue in the first half, up from about $1.2B a year earlier. most of the loss is accounting charges, not cash going out the door.

the line i keep rereading: up to $120.5B going to spacex for compute.

so part of every claude bill ends up at elon's company. wild timeline.