wcoenen

The article is about what LLMs can do when generating SVG images, not bitmaps. I think the reason for that is that LLMs can generate the XML text for SVG directly. The implications for AI safety are different, because there might not be any code involved that we can examine. So LLMs could start exchanging hidden messages without humans knowing about it.

But now that I think about it that way, it would probably be more interesting to see what kind of steganography LLMs would use in ordinary text. There's no reason to involve images.

> Heat pumps are mostly useless when it is minus -20 C.

Two-thirds of households in Norway are heated by heat-pump, so I don't think that can be true.

Nuclear plants can run with those hyperboloid cooling towers as the sink for waste heat. They are just a lot more expensive to build, run and maintain compared to using river water.

This post does not mention safety at all. What's to stop bad actors from fine tuning open weights to run fully automated genius-level scams personally targeting basically everybody?

I wonder what would happen if you give the model access to edit the conversation history itself? Would it try to fix the "glitches"?

> I don't see it happening.

Isn't it already happening with Cerebras? It's mentioned at the end of OpenAI's GPT 5.6 announcement:

"We're also launching GPT‑5.6 Sol on Cerebras at up to 750 tokens per second in July"

> why the EU doesn’t develop something like RuPay

We used to have Europay, which competed with Visa and MasterCard. But it merged with MasterCard in 2002.

The UX problem is elsewhere I think. Many users probably don't realize that the agent's context window is limited, and that clever compaction is happening regularly to make it seem infinite. But that necessarily means the agent has to forget stuff.

As a result, users will keep reusing the same coding or chat session again and again. While it would be better to start fresh for unrelated tasks.

You don't even need to double the tokens. Tokens are mapped to vectors right at the input of the LLM, so one of the numbers in that vector could be reserved to represent something like "authority". This way information about the source of each individual token can be injected right at the input.

System prompt tokens would get the maximum authority value, and random downloaded data would get the minimum authority value. Tokens from the user prompt could be somewhere in between.

Then train the model with examples that show that system prompts should be respected, and prompt injection attacks should be ignored.