Mohith kumar

[Mohith kumar @mohithkumar808]

POV: you send a 700k token prompt to an AI

and the model reads leftover memory from someone else's request

cloudflare says that hit 155 responses on its GLM-5.3 model, aug 27 to sep 9. an SGLang bug + a context window bigger than its cache

on sglang? upgrade to 0.5.22

[Mohith kumar @mohithkumar808]

what it doesn't mean

P vs NP is not solved
your encryption is fine
the method is still slow in practice

but if it holds, a lot of research built on ETH needs a second look. 700+ points on hacker news overnight for a reason

[Mohith kumar @mohithkumar808]

openai says its AI just proved 3 things mathematicians and physicists have chased for decades

including that a famous computer science assumption is false

here's what was claimed, and what's actually been checked 🧵

[Mohith kumar @mohithkumar808]

what's missing

which model did it: not said
how much compute: not said
independent sign-off: 0 of 3 so far

ETH's original authors, impagliazzo and paturi, plus ryan williams are now reviewing the CS claim

[Mohith kumar @mohithkumar808]

yesterday i posted that openai is cutting deals to win anthropic customers

now it cut the price for everyone

gpt-6.1 terra: $1 in / $5 out per 1M tokens, half of sol

per openai: 3x faster than sol, 50.3% on terminal-bench 4.0 (old terra 21.5%)

the price war is public now

[Mohith kumar @mohithkumar808]

80% less time on patient intake

thats what a clinic reports after an AI agent started asking the questions over text before the visit, per concurrence

same questions, just asked like a conversation

builders running intake or onboarding forms, whats your finish rate?

[Mohith kumar @mohithkumar808]

POV: you just finished patching PHP2Shell

then node and python get one too

broccoli2shell, CVE-2026-111497, CVSS 9.2, in google's brotli decoder. reported by openai's codex security

node 22.23.4 / 24.22.1 / 26.12.1
python Brotli 1.2.1
then restart everything

[Mohith kumar @mohithkumar808]

@AndrewCurran_ claude as the only author is wild. 14M output tokens and the acknowledgements section is one line lol

[Mohith kumar @mohithkumar808]

@murtuza_merc advisors doing tax paperwork by hand in 2026 is wild. no launch date tho so the spreadsheets live another quarter lol

[Mohith kumar @mohithkumar808]

@gregisenberg forms already went this way. the builder is free everywhere, the paid part is whatever chases people who quit on question 4

[Mohith kumar @mohithkumar808]

@IndependentEco @Compilethings free until the IPO filing makes everyone look at the margins lol. still, 16% to 52% is a lot of switching

[Mohith kumar @mohithkumar808]

yesterday i posted that replit moved its AI spend to openai over a discount

turns out it wasn't one deal

rippling: openai's share of its coding work went 16% to 52%, anthropic's fell 57% to 18%
datacamp: costs down 80 to 90%

per the information. the price war is here

[Mohith kumar @mohithkumar808]

@meaynx nothing on weights or a paper yet that i saw. would be wild if it ships open

[Mohith kumar @mohithkumar808]

block just showed an AI that learned to talk without ever reading a word

no text. no pretrained model. only recordings of people talking

and it runs on a macbook 🧵

[Mohith kumar @mohithkumar808]

@JatinChanchlan 2 people vs hundreds and the free tier did the selling. unlimited responses on free vs paying per response is a crazy wedge

[Mohith kumar @mohithkumar808]

@IndependentEco day 2 is rough. a 3am usage reset is the most dev alarm clock ever

[Mohith kumar @mohithkumar808]

@kimmonismus the clipboard in the waiting room finally dying lol. people actually finish it when its a conversation not 3 pages of boxes

[Mohith kumar @mohithkumar808]

if you sell AI to normal companies, this chart is the objection you'll hear all of 2027

what number would you show a CFO to prove AI paid off?

[Mohith kumar @mohithkumar808]

my take: margins are a lagging signal

most big companies bolted a chatbot onto the old process. the payoff shows up when the process itself changes, and that takes years, not quarters

but "we're early" can't be the answer forever