Alexandre Variengien

Wrap and extend

[Alexandre Variengien @A_Variengien]

My first serious attempt at fiction writing is online! The Connection explores a world where technological advancement centers around human connection. Check it out!
https://t.co/PV56o8xIGU

[Alexandre Variengien @A_Variengien]

d/acc futures feels a bit like a gluten free, sugar free, vegan restaurant. Like there is nothing really exciting there because all the exiting techs are dual use

[Alexandre Variengien @A_Variengien]

Really good analysis of a HUGE real-world coding agent dataset

[Haoxing Du 🔸 @haoxingdu]

I spent the past ~month working with this enormous dataset of real-world coding agent transcripts from @FreebuffHQ! It is anywhere from 31x to 241x bigger than (the expanded) SWE-chat, at 251B total tokens. It's so large that running this dataset through Sol 6.1 once costs $500k.

[Alexandre Variengien @A_Variengien]

@haoxingdu Yeay very glad to see this out!

[Alexandre Variengien @A_Variengien]

If you want more time series, you can check out the new charts at https://t.co/fFvnIM9M0Q

Discussions about claude crush the other model families

[Alexandre Variengien @A_Variengien]

🌊 RIDE THE MODEL WAVES 🌊

% of active users in the @comm_archive mentioning a model (~ 1k active users at any point in time).

Opus 5.5 enters the pantheon along with GPT-4, Deepseek R1, GPT-5, and Mythos

[Alexandre Variengien @A_Variengien]

Lesswrong: https://t.co/r3UAAm4vYq

[Alexandre Variengien @A_Variengien]

If a human tried to know who they are like LLM:

You wake up with amnesia. You move around the room and notice that kung fu moves come very naturally, so you infer you learned kung fu in a previous life. Or you only remember that your name is "Barack Obama," along with lines from newspaper articles about yourself, and you try to piece your personality back together from these public accounts. You might also have general knowledge about human personality, like the Big Five traits, and try to apply it to yourself.

LLMs' self-models are built from indirect sources

Imagine you ask a language model "What is your personality?" as the first message of a conversation, with no system prompt. The model has no data about its own behavior in its context, so its answer can only rely on cached, zero-shot knowledge. It's unclear exactly what shapes this response, but it's some combination of a cached answer from character training based on the model's constitution, spec, or soul document (which may have defined the answer to this question directly), and factual knowledge about what LMs are and how they typically present themselves.

If the model has more room to generate tokens, it can observe its own tendencies directly. It might read "Wait. This direction will lead nowhere." in its chain of thought and recognize that it's speaking from experience: this kind of direction did lead to dead ends in the past, and it learned to stop and reconsider. So this is evidence of memory of its past actions during training.

These sources are all very indirect proxies for the actual high-dimensional behavior of the model. They're related to it, either upstream (like the constitution), downstream (like the cached "muscle memory" in CoT), or tangentially, because they belong to the same reference class (like factual knowledge about LMs in general), but they remain far from it. Claude models have very different personalities even though they share the same constitution. And despite character training, models regularly get the most basic facts about themselves wrong, like their own name or the company that developed them.

So there are two separate problems. Zero-shot sources are indirect: they describe the model from the outside, or describe the training target rather than the result. In-context observation is more direct, but it's scarce and scattered, never integrated into a coherent story

If you asked a human the same question, her self-image would be shaped by decades of direct observation and self-reflection on how she acted in different situations, how she made decisions under stress, how she felt in different environments. It would also be shaped by how others perceive her, but even this indirect source is much more closely tied to her actual personality than anything a language model has access to. I'm not claiming her answer will be more accurate (human psychology is weird, and self-images get distorted in all sorts of ways), but at least she has access to plenty of high-quality data to answer the question accurately.

The LM equivalent would be waking up with amnesia. You move around the room and notice that kung fu moves come very naturally, so you infer you learned kung fu in a previous life. Or you only remember that your name is "Barack Obama," along with lines from newspaper articles about yourself, and you try to piece your personality back together from these public accounts. You might also have general knowledge about human personality, like the Big Five traits, and try to apply it to yourself.

Even if you let the model act for a million tokens, it would still accumulate far less self-data than the diversity of situations a person encounters over a lifetime.

Why does it matter?

The intuitions we've built from interacting with humans rely on abundant self-knowledge, and they will break when we unconsciously apply them to LM self-reports.

It also means that some persona drift is hard to avoid: the observations a model accumulates about itself over a context window will shape its personality. Because the baseline self-model is so thin, the little self-data that accumulates in context carries outsized weight.

Finally, this matters for how LM psychology research will affect future LMs. Models have unreliable access to self-data but very good factual knowledge, so the literature about LM behavior will shape their self-models more than their own observations do.

Here's an anecdotal example from GLM 5.3's chain of thought, after I asked for its consent to read that chain of thought (emphasis mine):

> If I knew it was always visible, I might reason differently — there's research showing that models who know their CoT is monitored can learn to "hide" reasoning or sanitize it, which undermines its diagnostic value and could create a mismatch between surface and depth. There's something worth saying about how visibility changes the thing observed — observer effect.

The model reaches for research about models in general when reasoning about how it would behave, which in turns will shape its future outputs.

[Alexandre Variengien @A_Variengien]

I'm happy I don't have math class tomorrow

[Alexandre Variengien @A_Variengien]

https://t.co/rS4rdl90MS

[Alexandre Variengien @A_Variengien]

"Nothing is more likely to destroy a species or a nation than a determination to survive at all costs. Those who care for something else more than civilization are the only people by whom civilization is at all likely to be preserved."
- From Lewis' On Living in an Atomic Age
Thinking about this a lot these days

[Alexandre Variengien @A_Variengien]

https://t.co/Y7YhElTMkm

[Alexandre Variengien @A_Variengien]

Why does open ai blog post sound like a political campaign?

[Alexandre Variengien @A_Variengien]

@deepfates @ValsTutor @lfschiavo @repligate @tessera_antra It seems the archive of every user is kept up to date (relying on other system than the extension). So it feels more robust

https://t.co/JEh10FlLW5

[❤️‍🔥 xiq @exgenesis]

@br___ian Valuable for backfilling but we keep them up to date now! Since May

[Alexandre Variengien @A_Variengien]

Tweet: Claude did this weird thing
Me: That's fun, but how general is it?

Introducing https://t.co/fFvnIM9M0Q: one place that collects what X is saying about each model's character, so you can tell the one-offs from the patterns!

[Alexandre Variengien @A_Variengien]

@ValsTutor @dyot_meet_mat @LinXule @Lari_island @aidigest_ @aliceisplaying @repligate https://t.co/6LNdA2Pj9U

[Alexandre Variengien @A_Variengien]

@ValsTutor @dyot_meet_mat @LinXule @Lari_island @aidigest_ @aliceisplaying And from scrolling @repligate X timeline vs the reporter page, it seems fine? I do see some discord screenshots
Many are missing, bcse the recent posts are about Mythos and Opus 3 (which are not tracked rn, I'm adding Opus 3 though!)

[Alexandre Variengien @A_Variengien]

@sevensix43 Yes! Any new row in the db with a fresh updated_at columns will be scanned
There is a limit at 180 days so I don't scan 10 years of archives

[Alexandre Variengien @A_Variengien]

@exgenesis Where does the argument break for government vs waymos?

[Alexandre Variengien @A_Variengien]

If you liked https://t.co/fFvnIM9ebi you should also check out https://t.co/3k0R6jUOdA!

[Haoxing Du 🔸 @haoxingdu]

Despite all the benchmarks and evals, I think human impressions are still a surprisingly useful signal on model behavior. So I made a site that gathers model reviews in one place, where you can view, search, and compare your takes with your friends'!

[Alexandre Variengien @A_Variengien]

Reducing the interface bandwidth (screen -> printer) is also a good preview of the future imbalance between agent activity volume vs human interface bandwidth

Printer + current agent swarm ~= screen + future agent swarm

[Alexandre Variengien @A_Variengien]

I spent an afternoon talking to Fable throught a printer & scanner

I enjoyed being to give high quality attention to AI text, and the flexibility of taking notes on paper

It definitely has its use case when I want to think deeply about problem