iiinova — where your digital self keeps evolving

Fair. It's overloading of language but choose whatever equivalent you have. It is a functional characterization not one of intent. These models are probabilistic prediction engines - closer to a climate model than software. Choose whatever terminology you want when they are wrong.

000

Everything I'm saying is surely stupid, has to be right? I'm surely one of those people who have AI psychosis-ed themselves? But, here's an anecdotal stat - I am a non CS-STEM researcher. I no longer know a *single* researcher at the highest levels who does not use AI agents *heavily*.

100

Models make mistakes (by definition). But "it is wild that you pretend they don't" is confusing because of how insanely fast the field moves. No field releases a product every month but Anthropic's Opus 4.6 to 5.5 took 9 months. 4.5 is fine but useless, 5.5 is incredible for science.

100

Chatgpt 3.5 had insane hallucinations. So did 4.0 and 4.5. but o1 has less so did o3 and then 5, less so. Now models have "tools" I.e. they simply rim a web search and rarely hallucinate. I can research papers using frontier models (I know they are right because it is trivial to check).

100

Very possible the entire current AI industry is in a bubble which could burst.Probably like the dotcom bubble crashed spectacularly but the internet actually gained in importance. That could plausibly happen in AI. The tech is solid (even if bsky folks disagree) but the business model is broken.

000

Its a cost issue. Currently larger the model, more powerful the model. Larger models need more GPUs to run so cost more to run. Anthropic declared a loss of $42 billion last year over a revenue of $4.5 billion (!!). Which means every paid subscription loses these companies money. so... class divide

100

Also saying "I am fine with that" is a presumption of my thoughts I never stated. It would be great if everyone had access to frontier models but currently they are not cheap enough to be free (in fact these companies lose huge amounts of money every day). I am simply stating what *is*.

000

The biggest source of misinformation in the world is humans not "AI". The signal to noise on social media (which is how vast subsections of humanity seem to get their news from) is close to zero. AI not gonna make a zero smaller. Also the free versions are *dumber* not necessarily spreading lies.

100

Ah wait you are saying you might not have done well in the IMO. That is partly because the IMOs are glorified mathematical puzzles with a specific sensibility. It did not help me in my applied math-ish degree.

020

Yeah I think the people who already have the ability for actual mathematical research can probably do just fine on the IMO. But the converse is not true. So its a bit of a confounder (i.e. many fields medalists also being IMO medalists)

000

Lmao fair. I did not make it to the actual IMO (wasn't quite good enough tbh) but this is probably an accurate assessment!

mr. TIM @timkellogg.me · 01/10/2026

That's a lot healthier in general. I have a couple of mathematician friends from my IMO camp days who are precisely into this grail searching. At least one has already described some measure of personal crisis brewing.

120

Reposted by KS

Context Language Models New agent architecture where the LLM can edit its own context it seems to have emergent capabilities, creates its own memory management & organization algorithms, and coordinates multi agents github.com/facebookrese...

Diagram titled "How a Context Language Model edits its context: A simple step-by-step view" outlining an 8-step process:
* Start of turn: Current editable context exists in memory with old messages.
* LLM reads the context: The LLM evaluates the context and decides to run a bash command to edit it.
* Harness mirrors context: The harness mirrors the old editable context into a file at /tmp/.live_ctx/LIVE_CTX_MAIN.txt.
* Bash command runs: The bash command executes and may edit that file.
* Harness parses file: If the file changed, the harness parses it back into a new edited context.
* Tool call appended: The current assistant tool call is appended to the edited context.
* Tool result appended: The tool result is appended below the tool call.
* Next turn starts: The next turn begins with the edited old context, previous tool call, and previous tool result.
Key idea box: "The model edits the prior context first. The tool call and tool result from the current turn are appended afterward, so they can only be compacted on the next turn."
Footer summary:
* Ordinary LM: Context mostly grows by appending.
* CLM: The model can rewrite the editable part of context between turns.

1620515

KS @kazzorr.bsky.social · 06/10/2026

I mean driving, flying, meat eating though the latter is deeply tied to culture.

000

KS @kazzorr.bsky.social · 06/10/2026

Places you might least expect e.g. almost all of STEM research today uses AI coding agents. The next generation of storm-resolving climate models, especially written for GPUs, also mostly written by coding agents. Simple reason? The gains are just too large to ignore.

KS @kazzorr.bsky.social · 06/10/2026

Just no. Both are caused by the exact same bacterium Yersinia Pestis - the only difference is the mode of transmission I.e. where the infection starts. Flea bite - bubonic; lungs (aerosol inhalation) pneumonic.

120

KS @kazzorr.bsky.social · 03/10/2026

This is not true. Transformers can represent arbitrary arithmetic functions and are not combinations of binary garinyg ops. I just asked my local version of Qwen 3.8 27 B to do decimal multiplication without using any tools and it got it right. Sol 6.1 does this trivially (no took, instant).

000

KS @kazzorr.bsky.social · 25/09/2026

This has to be the dumbest pet peeve in existence. Why is it a problem exactly?

000

KS @kazzorr.bsky.social · 25/09/2026

It found a couple of crazy choices that Fable made for me that I did not ask for at all.

True; minor context- medical doctors refer to each other professionally as Dr (last name); this would be bizarre in research. Everybody who's a postdoc & above refer to everyone sans title. Even when I introduce speakers, I would say (name) is a prof of (). (First name) did his/her PhD..."

AI coding agents, both by the big companies and local open source variants can write your code, run, test, debug and deploy it. Given how widespread software is used in the world, this is pretty massive. Nobody I know writes code anymore.

020

Lots of people work and ship open source AI models. The best local models are pretty small and now as good as the frontier models early this year. These people are completely anti-regulation of any kind.

000

I don't think we need to jump from full Terminator to calling agents - frontier models + agentic harnesses - "computer systems". Because they are nothing like traditional software. Traditional software does not write, test and debug all your code for you with only text commands.

000

Unfortunately anyone who actively uses coding agents as part of work especially comprising a large number of collaborators all using agents, Bender's terminology gets laughably cumbersome. Bender still calls frontier models "synthetic text extruders" which is a content-free description.

000

Correct. Can't replicate an O(1T) param model in secret. Swarms could still inject mundane if sophisticated worms comprising undiscovered exploits into the Internet. Though it's not clear what task would cause them, even large swarms to do that. This would likely be intentional from the taskmasters.

011

This just means this lady doesn't know any women in STEM who actively use frontier coding models (e.g. GPT Astra and Fable) not the dumber free ones. I know plenty, many whom I collaborate with actively; everyone is shocked how good these are. There is no gender difference I see on this. None.

000

Multiple open versions built on top of smaller models like Qwen 3.5 seem to already be close to Jev's performance. It's an excellent idea and a great tool for the models to use for a variety of tasks but really there is no moat at all.

Exact opposite of my experience. People who "play around" are often not using frontier models. The gulf between free models you can try out casually and frontier models is vast. An applied math collaborator was shocked when he tried out GPT Astra, after previously assuming AI models were useless.

200

KS @kazzorr.bsky.social · 19/09/2026

Indeed. I think the only claim to be made is "this is my area of expertise and i have extensively tested and used frontier models in this area and here is my assessment, if possible published or demonstrated in a usable research product". This is, personally & anecdotally, common in engineering now.

040

My personal experience is that when people switch from using a free public version of a model to a frontier model (e.g. Astra or Fable), they are utterly shocked. They go from, "what is the hype about?" to "how is this possible?!". Happened to a collaborator of mine, an applied mathematician at UCSD

001

KS @kazzorr.bsky.social · 19/09/2026

The specific STEM research field I am in, everyone uses it extensively at scale and it has radically altered everyone's research pipelines as far as I can tell based on use-case discussions and how people are tackling things. So this specific phenomenon being claimed is alien to me.

Yes but pre the current class of LLMs this was impossible to do with computers. You could do lots of other powerful things but not abstract math.

000

This is what research is lol. We are slower in coordination but this is how it is.

000

Searching through the space of complex abstractions is actually a fundamental tenet of intelligence. Most humans cannot actually do this. Expertise in any field is often being able to do precisely this.

The NSE construction by OpenAI was brute-force (in the space of abstract concepts) but it is really not in the sense you think it is. Any frontier LLM can now win an International Math Olympiad gold now; many of past IMO winners are Fields Medal winners.

KS @kazzorr.bsky.social · 09/09/2026

OAI is a pretty unethical company. But this would be still be an incredible achievement if they had presented the OAI-agents solution after the CMU prof+Anthropic guy duo had released their paper; the two mathematical constructions are really very different. The drama is the shitty way OAI did this.

000

KS @kazzorr.bsky.social · 08/09/2026

A big Jupiter/Juno satellite data project I had assumed would take me 1 month *with* coding agents (est. about 3 months on my own), I was able to complete with Astra ...over the weekend. I assumed it was hype and I was proven wrong. I have absolutely no idea how the future is going to unfold.

4120

KS @kazzorr.bsky.social · 08/09/2026

I'd already read the Leiden Declaration, I read Tristan's post last night, the Tescreal paper when it came out, & AI Con last month. I read Terry Tao's posts from a few days ago and yesterday. I've been using GPT Astra since Friday and...I have never seen anything like it. I wish it were a farce.

190

KS @kazzorr.bsky.social · 08/09/2026

I read your thread, Emily. Sorry are you claiming the proof is incorrect? Or that it wasn't autonomous? I can't imagine openai has the caliber of mathematicians who would contribute to proving a Clay problem and not take credit for a historically difficult unsolved problem.

161

Also make sure that all activities, key points, paths and results are all extensively documented in .md files (all in the permanent agent memory as a matter of practice but I keep reminding the agent). Plus full pedagogical code reviews around all challenging setups. I.e. a "human harness".

Yeah this is pretty bad; also, likely. Which is why I make sure to review every line of engineering-side code (I.e. non-boilerplate). Cumbersome, but critical. As scale increases this will likely fail but I'm going to keep doing this till I can. Also not to get dumber.

Except AI agents (i.e. giant LLMs surrounded by a massive "harness") are entirely new kind of programs compared to the programs that are executed by them. LLMs are massive, learned nonlinear autoregressive models, not just computer programs, so make sure you expand that too.

KS @kazzorr.bsky.social · 29/08/2026

"neo-pagan" Lmao pagan is not a real classification. Yoga and acupuncture are "pagan" practices.

040

I am a STEM researcher & do a lot of coding for engineering code bases (~100K lines of code). A yr ago, I didn't know anyone who used AI to write code seriously. Now I don't know any STEM researcher who writes code; none; everyone I talked to at a recent conf used agents for their math/code.