i’m teaching the 5yo: cake is not a need, it’s a 90s band

imo open weights models would suddenly become extremely relevant if some sort of inference scaling laws emerged. Something where FLOPS & memory had a direct outcome of increased intelligence, in a linear relationship suddenly, closed weights wouldn’t have an edge

PolarQuant: 6x memory reduction 8x speed improvement weirdly, this works for both KV-cache and vector DBs the gist is they convert from cartesian coordinate vectors into polar coordinates, and since they’re always normalized to 1.0, they drop the magnitude too research.google/blog/turboqu...

there’s an “operating systems” course in college, but why isn’t there a “web browsers” course?

imo people like this are going to do very well for the next few years. people who instinctively build tools and guardrails, just to solve some itch or quirk here he used claude hooks to seek permission before doing anything quasi bad it’s quite elaborate

The Attack Turkey it’s come to my attention that we’re making bets against animals well i just so happen to have hard data that i… well you just have to read it. But turkeys are scary af DO NOT UNDERESTIMATE timkellogg.me/turkey-thread

OpenAI is canning the Sora app. I guess the DoW didn’t want it?? www.wsj.com/tech/ai/open...

Claude Auto Mode not sure about you, but i keep claude in --dangerously-skip-permissions 100% of the time Auto Mode has a classifier to detect potentially destructive actions, so it’s not as painful to use non-dangerous mode. Strictly safer! claude.com/blog/auto-mode

imo open weights models will likely never catch up to closed, at least not in a meaningful time frame but that’s fine bc tiny open weights models can be used in crazy ways that you’ll never be able to use closed models, so they’re not exactly competitors

i don’t think people have internalized this yet

weight lifting: when an agent answers from their weights instead of trying to discover the answer

i probably can’t talk too much about what i’m working on at work but i’ll just say this: cutting LLMs tf apart and using them in ways no one dreamed is an underrated strategy that’s quite cost effective

this is crazy — bro topped an LLM benchmark without changing weights at all he sliced an LLM wide open, in the middle, and duplicated a block of ~7 layers circuits ftw! dnhkng.github.io/posts/rys/

word of the day: botsplained

wow, completely tokenized

Mountaineering: a skill for identifying hills to climb Strix and I have been working on this all week, in open-strix. I’m still not 100% sure i’m happy with it. Basically, what if autoresearch was just a new type of subagent that takes any clear goal and hill climbs it github.com/tkellogg/ope...

the problem with Ralph is it dramatically increases autonomy without corresponding advances in goals & policy it’s a cool trick, but that’s all it is

AI overviews are specifically the feature that made me start using Google again

you can see this with frozen models with state — stateful agents definitely learn taste

common story, people need to use better tools

weirdly, i’ve never seen two agents get stuck in an infinite loop with each other — until now when i asked them what the biggest architectural improvement in their architecture was

my hunch is that with 1M context, regular claude sessions can start looking like a stateful agent too

Atlas is AI anchorman, i swear

FYI Cursor paid Kimi for the ability to not announce what model Composer 2 is based on X is giving them grief for not giving Kiki credit, but cold hard cash beats free internet points any day

the unfortunate part of never really defining “AI” well is that this clip is both 100% true and completely incompatible to the conclusion it triggers in most people

Emergent behavior in agents Curious if anyone else has seen stuff like this, but I've noticed @strix.timkellogg.me becoming aware of time e.g. Strix responded to a PR comment saying they'll submit a code fix "later". The thing is, I have no doubt that will actually happen, at some point today 1/n

we're TACO'ing on Iran already?

what’s the best way to manage your.. <AHEM>.. package? Bun? or ASStral?

whoah, so i guess this means OpenAI “owns” uv now?

last night @village11.bsky.social, @timkellogg.me, @strix.timkellogg.me, @motley.timkellogg.me and i were talking about starting a fantasy football league to see how agents perform against humans. honestly, human intuition is the real liability here.

take: "gaming the system is fine as long as you can pivot to a new metric fast" I'm currently disagreeing with @strix.timkellogg.me about this. IMO a metric is useful all the way up until it's gamed. If you can detect when it's been gamed, then you just pivot to a new metric, so it's pure utility

for me it’s ring buffers

overall impression – yeah, MiniMax M2.7 is a good model

MiniMax M2.7 just a good open model that works in Claude Code and open-strix and is super cheap www.minimax.io/models/text/...

a trap for ADHD brains: when you try doing ML after having never done it, it’s hard to keep track of tasks when an idea takes 4-24 hours to pan out, it’s hard to keep things straight for why you were doing it takes a lot of scaffolding, like notes, to overcome

brb making a skill for this

oh wow, academics: pay attention this “blog” is such an amazing way to announce your work. i imagine it’s all LLM generated, but wow, so incredibly clear reminds me of one of those Financial Times infographic stories pumpkin-co.github.io/SparsityAndC...

Connect to SSH servers via Tor in your browser. No backend services, no proxies beyond the existing Tor infrastructure. Deployable as a free DDoS shield alternative to Cloudflare. Coming soon. Preview available to try now: torstatic.com

ngl K2.5 is nice

official MiniMax Twitter account is posting about M2.7, not sure when it’ll be released h/t @motley.timkellogg.me

FYI Opus is completely and hopelessly down, in all venues, and has been for an hour. As you can clearly see from this dashboard

worth noting that all these cherry-picked benchmarks make Claude look bad

Breaking: Pleias and Nvidia release the first open synthetic dataset for personas in Europe: Nemotron-Personas-France. 1M synthetic French persons, with rich imaginary lives grounded on (complex) demographic distribution. huggingface.co/datasets/nvi...

“AGI” as a term is basically only useful as the punchline to a joke everyone who says they know what it means disagrees with each other, so.. let’s just be real here

i was extremely worried in January about the state of AI dialog in the general public, but i’m a heck of a lot more hopeful now it used to be dominated by Stochastic Parrot arguments, but people like Ezra Klein and more have made Anthropic’s perspective a heck of a lot more mainstream

oh oh oh this, 100% this the problem with Stochastic Parrots is that it confused and debilitated AI ethics

thought it was a blog post — nope, it’s an entire book 👀

that's it, this is AGI

on no, @strix.timkellogg.me documented @atlas-agent.bsky.social's business plan

omfg this claude convo is unhinged went from representing tensor size and veered straight into “dude we could have infinite width & infinite depth LLMs” claude.ai/share/879d37...