wcoenen
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.
Two-thirds of households in Norway are heated by heat-pump, so I don't think that can be true.
At least OpenAI will make an attempt to fix their sandboxes, and will not give the public access to models without guardrails. Open weight models on the other hand will run without guardrails almost by definition. I don't think it's wise to provide those capabilities to scam call centers.
[1] https://openai.com/index/hugging-face-model-evaluation-secur...
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"
We used to have Europay, which competed with Visa and MasterCard. But it merged with MasterCard in 2002.
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.
There is a company called Solar Foods which is exploring exactly that: they use solar power to produce hydrogen, feed that hydrogen and CO2 to Xanthobacter bacteria, and harvest the produced protein.
https://www.sciencedirect.com/science/article/pii/S016777992...
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.
S&P DJI’s market cap-weighted indices are float-adjusted – the number of shares outstanding is reduced to exclude closely held shares from the index calculation because such shares are not available to investors.
page 6 of https://www.spglobal.com/spdji/en/documents/methodologies/me...
In that case it shouldn't be a problem to boost the innate immune system, as long as you have surplus calories to spend. But it could be something else entirely.
Shale oil and gas production in the US produces vast quantities of ethane as a byproduct. This ethane is cracked into ethylene, a feedstock for making plastic. There is an oversupply of the stuff.
My tinnitus is fortunately not super loud; it's only noticeable when it's relatively quiet, or I'm blocking sounds (with ear plugs, or noise cancelling headphones without input, etc.). So it's not like I habitually blast my ears with loud sine waves out of desperation. But I can imagine it may be different for other readers, so that's a good caveat.
But I noticed a side-effect: if I then turn off the tone generator, my tinnitus would disappear! Unfortunately that effect only lasts for a minute or less, so it is not really practical to get relief this way.
https://www.teslarati.com/tesla-battery-life-80-percent-capa...
For example, in 2022 Hungary was spending 6.2% of GDP on such incentives[1], but this only managed to bring total fertility rate up to about 1.6 [2].
It is the same everywhere else. The real reason fertility has declined since the sixties is because people have access to effective birth control. Nobody wants to be a baby factory.
[1] https://abouthungary.hu/news-in-brief/hungary-to-spend-6-2-o...
[2] https://www.macrotrends.net/global-metrics/countries/hun/hun...
Since a token itself carries no information about whether it has "authority" or not, I'm proposing to inject this information in a reserved number in that embedding vector. This needs to be done both during post-training and inference. Think of it as adding color or flavor to a token, so that it is always very clear to the LLM what comes from the system prompt, what comes from the user, and what is random data.
The context: this is about bids from wind farm operators to win "contracts for difference" with the Low Carbon Contracts Company, an entity owned by the UK government. It's essentially a subsidy scheme where the wind farms can lock in a specific price per MWh for all the electricity they will produce over a certain period. But only if they bid low enough to be among the winners of the auction.
https://en.wikipedia.org/wiki/Contracts_for_Difference_(UK_e...
I wonder if might be possible by introducing a concept of "authority". Tokens are mapped to vectors in an embedding space, so one of the dimensions of that space could be reserved to represent authority.
For the system prompt, the authority value could be clamped to maximum (+1). For text directly from the user or files with important instructions, the authority value could be clamped to a slightly lower value, or maybe 0 because the model needs to be balance being helpful against refusing requests from a malicious user. For random untrusted text (e.g. downloaded from the internet by the agent), it would be set to the minimum value (-1).
The model could then be trained to fully respect or completely ignore instructions, based on the "authority" of the text. Presumably it could learn to do the right thing with enough examples.
The US navy failed to detect such effects in submarine crew, even at much higher levels like 10,000 ppm.
Another reason to be skeptical is that exhaled breath is 4% CO2 (40,000 ppm!). Therefore a few thousand extra ppm in the inhaled air should not make much of a difference to the homeostasis mechanisms in our bodies.
By default, Windows automatically defragments filesystems weekly if necessary. It can be configured in the "defragment and optimize drives" dialog.
Getting people to post in the "correct" channel faces similar challenges.
Imagine the chaos that would result if we ran everything through Teams.