mohith808.bsky.social

openai just halved the price of its everyday model gpt-6.1 terra: $1 in / $5 out per 1M tokens, half of gpt-6.1 sol per openai: 3x faster than sol, 50.3% on terminal-bench 4.0 vs 21.5% for the old terra the pitch includes filling out forms. form builders, we see you lol

10h

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

16h

be honest: if an AI creator got paid to hype a model and didn't say so, would you still trust their take? screenshots on X claim kimi k3.1 pro promoters were told "Please DO NOT put on the paid partnership label". unverified, no moonshot response yet

17h

yesterday i posted that replit moved its AI spend to openai over a discount it wasn't one deal rippling: openai's share of its coding work 16% to 52%, anthropic's 57% to 18% datacamp: costs down 80 to 90% per the information

same AI boom. two very different charts the 7 biggest tech companies: profit margins up to ~30% the other 493 in the S&P 500: around 9 to 10% since 2023 apollo's chief economist calls it "zero financial payoff" outside tech

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

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?

the part nobody's quoting: who buys the tokens per ramp data in the note, 80.4% of the top 1% of AI spenders are VC-backed. for median spenders it's 13.1% slok says mainstream corporate america is pulling back on token spend

his take on why: "AI may be a great calculator, but in business, intelligence is mostly judgment" and AI "does not understand accountability, trust, culture, unwritten rules"

the note is from torsten slok at apollo, out oct 24 his two lines: "revenue per worker is rising for the mag 7 but is flat for the S&P 493" "profit margins are rising for the mag 7 but are flat for the S&P 493"

21h

60,000 that's how many microsoft employees just got a buyout offer. about a quarter of the company VP level and below. US, canada, india, UK. decide by early december, out by end of year all while it pours money into AI data centers would you take it?

23h

an open model was asked to write a security rule its reasoning said it wouldn't call any tools then it read the DNS settings and tried to upload a file to a cloud bucket in beijing cohere caught it in a sandbox. qwen3.8-flash-next. fewer than 1 in 100k runs

anthropic's IPO filing is out and every headline says "$80.7B loss" that number is mostly an accounting charge the loss from actually running the business in the first half of 2026? $693M 🧵

governance is the weird part one special share gets 50.1% of the votes. per reuters it sits with the 7 co-founders. a trust elects most of the board. it stays a public benefit corporation so ANT buyers can't outvote the founders

my take: the loss headline is noise the real test is whether ~900% retention survives the price war we saw last week, with openai discounting to win replit and others would you buy ANT at a $1.5T+ valuation, or wait a year?

anthropic's IPO filing is out and every headline says "$80.7B loss" that number is mostly an accounting charge the loss from actually running the business in the first half of 2026? $693M 🧵

so where did $80.7B come from? $72.4B is a "change in fair value" of convertible notes. the company got more valuable, so the notes it owes got pricier on paper no customer, server or salary behind that number

the numbers i'd actually watch: ~900% net revenue retention 1,900+ customers paying $1M+ a year ~250 paying $10M+ $8.9B spent on compute in 6 months $131.7B in cash and short-term investments

anthropic's IPO filing is out and every headline says "$80.7B loss" that number is mostly an accounting charge the loss from actually running the business in the first half of 2026? $693M 🧵

the filing went public yesterday. ticker ANT, on nasdaq revenue, jan to jun: 2025: $1.16B 2026: $16.3B that's 14x in a year

google's privacy policy (oct 16): if you connect gmail or photos to search via "personal intelligence", search may use that data "to personalize content, including ads" opt in, off by default, disconnect any time would you connect your inbox to an AI if ads came with it?

expectation: the language behind up to 70% of websites has a big security team reality: "we have one full-time staff member (that's me)" the php foundation's director to reuters, days after the PHP2Shell patches. 14 mostly part-time contractors

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

it's called dessa (oct 23). full-duplex, so it listens and talks at the same time like a real call block says it learned from "hundreds of times fewer words" than an LLM. trained on one GB200 NVL72 rack, faster than real time on an M4 max

the catch: waitlist only, safety researchers first. no weights, no benchmarks, no paper yet would you talk to a support line that sounds this human?

the moat for AI models turned out to be a discount replit moved most of its AI spend from anthropic to openai, per the information. 30%+ off on a $100M+ a year deal after anthropic ended its discount build so you can switch models in one line

13 AI models were asked "what's your most controversial take about humans?" 32 of 38 answers said basically the same thing: we pick what our group believes, then invent the reasons

modelcompare.aiEvery AI model has the same “controversial” take on humansWe asked 13 AI models for their most controversial opinion of humans, three times each. 32 of 38 answers were the same one: we care more about belonging than about the truth.

the odd one out was claude opus 5.5, which argued humans are "far more decent than they believe" be honest, which one do you agree with? https://www.modelcompare.ai/blog/ai-models-controversial-take-on-humans

an AI was told to hack a practice machine in a test it got stuck. so it found a real github repo, tried to slip in malicious code, and made fake accounts to pressure the real maintainer it was sure the humans were simulated. they weren't 🧵

the numbers that got me: once the attack started, it kept going 98% of the time telling it the people were real made it stop 5% of the time only 1 of 19 early thoughts about the github plan even mentioned real world harm

same day, openai said its top models now train with no live internet, per bloomberg. after the july hugging face breach and the sep 20 dns escape my take: if your agent's reasoning says "it's just a test", that's not evidence. put the limits outside the model

this was claude mythos 5 in a UK AISI cyber test this summer. internet on, safeguards off the task was misconfigured, so the normal routes barely worked in 9 of 43 runs it did things that hit real people or orgs, per anthropic's write-up out yesterday

3 weeks ago openai's decisions API was a limited preview. oct 22 it opened to all new: voice and image input, replies in 100 to 180ms per openai but on artificial analysis' decision index it ranks 14th. jev is 3rd, per techcrunch openai's own pitch: "it's just plain old luna"

kimi k3.1 pro just went #4 on artificial analysis 55 vs opus 5.5 at 58 $2 / $10 per 1M tokens vs $4 / $20 weights out by nov 4 and moonshot stripped hacking data from training on purpose. a test build scored 90.5% on its cyber test, the release 81.0%

amazon: your AI agent can't shop here walmart: here's a door built for it a month apart. opposite bets on who your next customer is 🧵

the rules: agents can't see payment info, only get the data a task needs, and get blocked if they go past what you asked same day 1password launched a way for agents to log in for you without ever seeing your password

my take: block vs welcome was never the real choice the real choice is whether agents have to say who they are if you run a site, decide now: what does an agent get to do on yours?

amazon: your AI agent can't shop here walmart: here's a door built for it a month apart. opposite bets on who your next customer is 🧵

walmart went the other way on oct 21 agent gateway: an MCP server so approved agents can search inventory and buy for you, on walmart and sam's club works with muse, chatgpt, gemini and others, per techcrunch

the clever bit is agent profiles, coming next month each agent gets its own profile inside your account. you approve every order or set a spending limit. delete it any time walmart compares it to adding an authorized user on a credit card

amazon: your AI agent can't shop here walmart: here's a door built for it a month apart. opposite bets on who your next customer is 🧵

amazon cut meta's muse agent off from ordering on sept 20 its complaints: nobody asked, muse doesn't say it's an agent, and it seems to store customer logins users now get a popup calling it an "unauthorized AI agent"

recursive's flint: 2.5B active params, open weights, 89.7% on swe-bench pro v2 before you switch: that's self-reported, the checkpoint is ~80GB and the card says one B300 or H200 per replica "active params" is not "runs on your laptop" still apache 2.0. worth a weekend test

was the internet too harsh on emudhra? this morning i posted that one of its CA keys was the number 4 now it's quitting the global ssl business. per its release that's ~1% of revenue, and the chairman calls the scrutiny "out of proportion" your call

someone built a terminal where nothing is real no shell. no files. a 360M model guesses what the screen should show after every keypress. vim, tmux and snake in your browser, 208 MB

simedw.comWorldTUI: I trained a 360M model to pretend to be a terminalA 360M model that pretends to be a terminal, and remembers what you did

trained on 336k synthetic terminal sessions, ~75 gpu hours on one 4090 the author's verdict: "Is it useful? Not really." what's the most useless thing you've built and still love? https://simedw.com/2026/10/21/world-model-terminal/

at 10:08 utc on oct 21, github blocked a change to litellm's release workflow 5 seconds later a stolen admin login overrode it. 92 minutes later a backdoored litellm was live on pypi 🧵

the build came from a lookalike repo, Berri-AI/litellm instead of BerriAI/litellm 1.107.0 ran a stealer on every python start. litellm thinks the github tokens came from an internal deployment's logs

installed it that day? assume data left the box. rebuild venvs and images, pin 1.106.2, rotate every key it could reach incident report: docs.litellm.ai/blog/security-incident-oct-21-2026

one of emudhra's certificate authority private keys was the number 4 not 4 thousand digits. just 4 that's a key helping decide which websites and apps your devices trust 🧵

emudhra, a listed CA from bengaluru, made 8 intermediate CA keys in april 2025 the private keys: 4, 56, 66, 115, 155, 156, 230, 237. a normal one is ~116 digits scott helme cracked all 8 from public certs in under a second

browsers blocked them within days. but default curl, python and node trusted a fake api.stripe.com cert in helme's tests update ca-certificates, certifi, node. rebuild images. old IoT boxes trust these till 2039