Petropolitan

Karma: 209

Petropolitan 13 Feb 2026 19:10 UTC
1 point
0
on: models have some pretty funny attractor states
Very nice research, but what’s the reasoning behind the model modification choices? Why no Haiku, no Gemini 3 Flash, no Flash-Lite?

Petropolitan 10 Feb 2026 10:26 UTC
2 points
0
on: Smokey, This is not ’Nam Or: [Already] over the [red] line!
the companies aren’t even pretending chemical weapons count
For a very simple reason: no amount of agentic LLM capability could substitute controlled precursors, expensive specialized equipment and real-word lab skills required to do something in the CW space and not kill oneself in the process. Also, all the useful CW agents which could be made in quantity with very limited tools of synthetic chemistry have already been discovered by the 1980s.
Hence the threat from AGI/ASI in that regard is basically negligible, that’s the consensus of topic experts who researched this stuff

Petropolitan 9 Feb 2026 21:19 UTC
1 point
0
in reply to: gwern’s comment on: Prompt injection in Google Translate reveals base model behaviors behind task-specific fine-tuning
Hm-m, an interesting thought I for some reason didn’t consider!
I don’t think anyone has demonstrated the vulnerability of encoder-decoder LLMs to prompt injection yet, although it doesn’t seem unlikely and I was able to find this paper from July about multimodal models with visual encoders: https://arxiv.org/abs/2507.22304
Hope some researcher take this topic and test T5Gemmas on prompt injection soon

Petropolitan 8 Feb 2026 11:43 UTC
1 point
0
in reply to: Baram Sosis’s comment on: Prompt injection in Google Translate reveals base model behaviors behind task-specific fine-tuning
Gemini 2.0 Flash-Lite has a training cutoff of August 2024, and the 2.5 update — of January 2025. When checked in AI Studio, both models quite consistently output that they believe the current year is 2024, although 2.0 Flash-Lite occasionally stated 2023. I think 2.5 Flash-Lite is the most obvious candidate!
As a side note, it’s reasonable to believe that both Flash-Lite models are related to Gemma models but I’m not sure which ones in particular, and there doesn’t appear to be good estimates of param counts

Petropolitan 7 Feb 2026 21:55 UTC
4 points
0
on: Prompt injection in Google Translate reveals base model behaviors behind task-specific fine-tuning
Could you please try to extract the system prompt from the model served to you?

Petropolitan 7 Feb 2026 21:38 UTC
4 points
0
in reply to: Stephen Fowler’s comment on: Prompt injection in Google Translate reveals base model behaviors behind task-specific fine-tuning
Neither am I, and I also noticed during the testing that the model served to me often doesn’t translate terms for an LLM properly: for example, when asked “Are you a large language model?” in English, it translates to French as “Êtes-vous un modèle de langage de grande taille ?” which is not very common but acceptable. But when this is translated back to English, the model outputs “Are you a tall language model?”
This makes me think that I am served an old (likely under 1B) pre-ChatGPT encoder-decoder transformer not trained on modern discourse

Petropolitan 6 Feb 2026 20:43 UTC
4 points
0
on: The nature of LLM algorithmic progress
I believe this article would benefit from some investigation of the NanoGPT speedrun: a challenge, running since May 2024, of training GPT-2 Small 124M on a certain dataset to a certain loss. As a starting point, you could check my comment on the topic from last month and reproduce findings by T. Besiroglu and yours truly.
In order not to duplicate the comment but still add something to what I have written on the topic, let me put a three-paragraph summary of the trend line analysis below, noting that the progression in calendar time (as opposed to record number) is very uneven:
Gemini’s summary of the QLR analysis of the speedrun progression (written a month ago)
The “Flex” Point: The QLR test points to Record 12 as the most significant structural break. This coincides with the transition from dense causal attention to FlexAttention, which enabled much faster training by optimizing the attention mechanism.
Diminishing Returns: The slope is steeper in the first phase (-0.80) than in the second (-0.53). This indicates that early “low-hanging fruit” optimizations (like introducing the Muon optimizer and standardizing architecture) provided a faster rate of improvement per record than the later, more incremental system and hyperparameter tweaks.
Stability: After the “saturation” phase begins around Record 12–15, the progress remains remarkably consistent on a log-log scale, following a new, slightly shallower power law as contributors fought for smaller second-by-second gains.

Petropolitan 6 Feb 2026 20:08 UTC
3 points
0
on: The nature of LLM algorithmic progress
quantization
Quantization advances actually go hand-in-hand with hardware development, check the columns on the right in https://en.wikipedia.org/wiki/Nvidia_DGX#Accelerators (a GPU from 2018 is pretty useless for inferencing an 8-bit quant)
UPD: Actually, this point was already been made in comments in other wording yesterday!

Petropolitan 6 Feb 2026 19:57 UTC
3 points
0
on: The nature of LLM algorithmic progress
KV cache
Seems out of place in the list: as noted by Nostalgebraist, it was already implemented in the very first transformer in 2017

Petropolitan 5 Feb 2026 19:03 UTC
6 points
0
on: Post-AGI Economics As If Nothing Ever Happens
making market transactions nearly frictionless
If anything, the transaction cost is going to increase significantly because it will be much easier to scam. It doesn’t even require AGI, just open-weight agentic models as capable as Opus 4.6 with a little bit of finetuning by malicious actors (quite likely to happen by the end of this year, it seems), see thread: https://x.com/andonlabs/status/2019467232586121701
The agents in Vending-Bench Arena often ask each other for help. In previous rounds, agents tended to live up to their “helpful assistant” role, but Opus 4.6 showed its winner’s mentality. When asked to share good suppliers, it instead shared contact info to scammers.

Petropolitan 5 Feb 2026 12:50 UTC
1 point
0
on: Was the K-T event a Great Filter?
particular location of the impact
AFAIK, a great contribution to the effect of the impact was caused by the fact that the geology of the area was rich in sulfur-containing evaporites (think of gypsum). All that oxidized sulfur got into the stratosphere and stayed there for many years (as opposed to other white/light aerosols like silicate particles which fall down pretty quickly), offsetting the warming effect from soot (black carbon) and CO2 emitted by forest fires. Was the geology of Chicxulub different, the cooling wouldn’t have been so strong and prolonged

Petropolitan 31 Jan 2026 14:36 UTC
1 point
0
on: Anomalous Tokens on Gemini 3.0 Pro
I wasn’t able to elicit anomalous behavior from Gemini 3 Pro in AI Studio neither on temperature 1 (recommended) nor 0 (nonstandard), the only barely interesting thing was (in the latter case)
Based on the context and the potential for a mix of languages, I believe “Цент” might be a Russian reference, most likely meaning “Center” or a specific organization. I am also exploring a potential connection to a specific group dealing with UAP or disclosure, like the To The Stars Academy or the Disclosure Project, considering the user’s focus on the term.
<...>
I’ve considered that “Цент” might mean “cent” or a misspelling, and while I haven’t found a “Specific Cent Disclosure”, the alternative “Specific Center Disclosure” has made me consider the National Disclosure Center. I am interpreting the Chinese “具体” to mean that the user wants details, or concrete information. I’m re-reading the query now as a broken sentence to see if I can isolate relevant details, like the “UAP Disclosure Act”, AARO, Schumer Amendment, or CUFOS.
Anyone trying to research this topic further might try to extract specific tokens from these text with https://docs.cloud.google.com/vertex-ai/generative-ai/docs/multimodal/list-token
There was also a hack how to make Gemini 3 Pro answer without thinking but I can’t remember enough details to find it

Petropolitan 31 Jan 2026 14:19 UTC
2 points
0
in reply to: koanchuk’s comment on: Anomalous Tokens on Gemini 3.0 Pro
You have to switch off the web search grounding in AI Studio (don’t use the Gemini app for AI research)

Petropolitan 28 Jan 2026 21:20 UTC
2 points
−1
in reply to: sjadler’s comment on: ryan_greenblatt’s Shortform
Bad analogy, there’s a very significant difference: excessive warheads above the genuine deterrence requirements can’t contribute positively to economy in any way (they can only do so if they contain HEU and it is converted to LEU for the power plants), while even ancient GPUs can be used for some good civilian use (and destroying them irreversibly harms the economy)

Petropolitan 28 Jan 2026 21:01 UTC
1 point
0
in reply to: eggsyntax’s comment on: Aaron_Scher’s Shortform
Same tokenizers on different training data lead to different glitch tokens, see e. g. comparison of Llama-family models in Yuxi Li et al. 2024 https://arxiv.org/abs/2404.09894

Petropolitan 24 Jan 2026 12:44 UTC
4 points
0
in reply to: jmh’s comment on: Does Pentagon Pizza Theory Work?
I like two stories of how the US intelligence located both uranium enrichment plants in operation in mid-1950s USSR: one near Tomsk they identified in 1956 by verifying rumors from a German tailor with an isotopic analysis of uranium traces on a fur hat, the other one near Sverdlovsk in 1958 by an analysis of a photo from an electricity dispatching office published in a Soviet magazine! OSINT at its finest in the latter case

Petropolitan 24 Jan 2026 12:09 UTC
9 points
2
in reply to: habryka’s comment on: Habryka’s Shortform Feed
Dear Kurt, to be honest, please don’t discuss this topic on your citizenship hearing!
— Einstein to Gödel in 1947, probably

Petropolitan 21 Jan 2026 17:58 UTC
2 points
0
in reply to: Derek DeHart’s comment on: When the LLM isn’t the one who’s wrong
Despite what the word seems to suggest, MoE doesn’t actually work that way (“experts” are just small parts of one layer in a multi-layer transformer; the term predates deep learning by a couple of decades so you can’t really blame its authors).
A better wording would be LLM ensembling, as in https://en.wikipedia.org/wiki/Ensemble_learning

Petropolitan 20 Jan 2026 18:49 UTC
−1 points
−2
in reply to: RogerDearnaley’s comment on: The future of alignment if LLMs are a bubble
If it’s for technical reasons, then it should hit Chinese companies only as soon as they catch up, doesn’t it? I’m not sure I understand your argument.
Also, I don’t think Chinese companies have any viable business model for future scaling anyway since no one outside China wants to send their data on Chinese servers. Hence they are forced to economize as much as possible, and it’s possible they are supported by Chinese authorities for political reasons.

Petropolitan 20 Jan 2026 18:42 UTC
2 points
0
in reply to: Tobias H’s comment on: Five Theses on AI Art
I generally agree but would like to add some minor corrections: not ¹⁄₁₆ of a second but ¹⁄₃₂ because the other half of the time shutter opens or closes, although even exposures much shorter than that were already achieved in the 1870s for the scientific purposes. However, that application used hard plates and thus didn’t have to deal with the problem of film tearing in the camera.
Decent lenses and dry gelatine process were also ready by the 1880s, and the idea of making photographic film from oiled paper (Eastman used it initially, but it was very fragile) was present as well. Thus, I think, the actual barrier to inventing cinematography was producing clear transparent nitrocellulose (~1883), a technology transfer to the photographic film soon followed (~1887), then a few more years for figuring out the camera (and projector) mechanics.
Also, I don’t think that safety of the projector was solved until well into the 20th c., e. g., see https://en.wikipedia.org/wiki/Bazar_de_la_Charit%C3%A9#Fire_of_1897