half of the internet is down today 🎉

Gemini 3 is live

faster and more secure btw

Thought signatures are encrypted representations of the model's internal thought process and are used to preserve reasoning context across multi-turn interactions. All responses include a thought_signature field. As a general rule, if you receive a thought signature in a model response, you should pass it back exactly as received when sending the conversation history in the next turn. Failure to circulate thought signatures may cause the response to fail. Check the thought signature documentation for more explanations of signatures overall.

https://ai.google.dev/gemini-api/docs/image-generation#thought-signatures

https://ai.google.dev/gemini-api/docs/thought-signatures

https://github.com/monsieurPale/RSA-Backdoor

#rsa #go

may be a stupid question but, why is there no context handling system that just discards tokens from processing if they are not deemed required so that it doesn't clutter with useless tokens?

The short answer is: we are trying to build exactly that, but it is incredibly difficult to define "useless" before it's too late.

In fact, there is a lot of active research into this (techniques like "Token Pruning," "Sparse Attention," or "Selective Context"), but they haven't become the standard yet for a few specific reasons.

### 1. The "Whodunit" Problem (Predicting Relevance)
The main issue is that a token's "usefulness" is not a static property. A word that seems like filler in sentence 1 might be the critical key to understanding a reference in sentence 50.
*   Dynamic Importance: To know if a token is useless, the model effectively has to "read" it and understand its relationship to the *future* response. By the time it has done the math to figure out "I don't need this," it has already spent the compute processing it.
*   Hindsight is 20/20: Systems that do this often have to guess. If they guess wrong and discard a token that turns out to be crucial later, the model creates a "hallucination" or loses coherence because the context is literally missing.

### 2. The "Attention Sink" Phenomenon
This is a really weird quirk of how Transformer models (the architecture behind GPT, Claude, Gemini) work mathematically.
*   Researchers found that models often dump a huge amount of "attention" score onto tokens that seem meaningless (like the very first token in a sentence, or random punctuation).
*   If you delete these "useless" tokens, the model's math breaks down and it starts spewing nonsense. These are called "Attention Sinks." The model essentially uses these tokens as a garbage dump for extra probability mass so it doesn't get confused. You can't just delete them without retraining the model to handle it.[1]

### 3. It Breaks the Flow
Language is a sequence.[1] Even "useless" filler words (like "um," "so," "basically") carry prosody and structural information that helps the model maintain a conversational rhythm.
*   If you aggressively strip these out, the model might turn into a keyword-generating robot rather than a fluent assistant.
*   LazyLLM & Sparse Attention: There are newer techniques (like *LazyLLM* or *H2O*) that try to dynamically keep only the "heavy hitters" (tokens that get high attention scores) and drop the rest. This is working in labs, but deploying it at scale without degrading quality is still a massive engineering hurdle.

### Summary
We essentially have to choose between speed/efficiency (pruning tokens) and intelligence/accuracy (keeping everything just in case). Currently, it's safer to just "eat" the cost of processing useless tokens than to risk throwing away the one key detail that changes the answer.

===
Citations:

[1] lesswrong.com

===
Model: Gemini 3 Pro (Preview) (high reasoning, with search)