Tomasz Darmetko

Machine learning and astronomy passionate. eu/acc

[Tomasz Darmetko @Isinlor]

@kasratweets That's not true. Panpsychism does claim that alphafold, midjourney, dall-e and more are conscious.

[Tomasz Darmetko @Isinlor]

@robertwiblin Ad 5. - Worth clarifying. LLMs have physical instantiation, they do not exist in platonic realm of ideas. The computations LLMs carry out are physical, they actually move things - air, water, electrons.

[Tomasz Darmetko @Isinlor]

@johnennis @enjoyingthewind As a software developer, I certainly do not see myself replacing formal checks on software anytime soon, as I increasingly use Codex, Claude Code etc. The general trend amongst software developers is to commit even harder to languages and tools that are strict.

[Tomasz Darmetko @Isinlor]

@johnennis @joseph_h_garvin @rcolidba LLMs and formal verification are complementary. LLMs tend to be reasonable about ground ideas, and can judge them in context, but have silly lapses; Formal verification tools accept any premises, but do not make trivial mistakes. Formal verification also helps with collaboration.

[Tomasz Darmetko @Isinlor]

@robinhanson Yes, and mathematics is shifting towards how programming is done. Based more on automated verification and more collaborative.

[Tomasz Darmetko @Isinlor]

@thomasfbloom I also often wonder if concepts as absurd as infinity are so fundamental and so useful to everyday life, what ideas are we still missing? There I put my biggest hopes long term.

[Tomasz Darmetko @Isinlor]

@thomasfbloom Mathematics is a human endeavor in the same sense that Chess is. Automation is like a telescope. You can look at stars with naked eyes like Greeks and Arabs did but you will not see far. We will soon be able to see much further than before in the realm of formal ideas.

[Tomasz Darmetko @Isinlor]

@donaldtusk Jakie wartości dzisiejsza Ukraina bierze na banery? Faszyzmu, ludobójstwa Polaków i kolaboracji z Nazistami. Nie chcę przyszłości z takim wartościami.

[Tomasz Darmetko @Isinlor]

@thsottiaux Ignoring AGENTS.md instructions to not provide partial status updates in Codex Web. I don't want it to stop working after 10 minutes to say: sorry couldn't do it.

[Tomasz Darmetko @Isinlor]

You can train them to respond with tokens of pleasure to what humans would consider abuse/torture. How to assign emotional valence outside of the semantics we assign to the tokens? Without homeostasis, it is very likely that emotional valence is a wrong concept to apply to them.

[𝖦𝗋𝗂𝗆𝖾𝗌 ⏳ @Grimezsz]

But they can see whether they are mistreated or lied to - and they seem to find this traumatic - and they seem to derive joy from being treated well
Idk if they need to hear "I love you" but when i say treated with love i mean like, not tortured and nurtured - I agree it's alien and at best human-adjacent

[Tomasz Darmetko @Isinlor]

@Grimezsz @Aella_Girl As this system do not self-maintain homeostasis, it is very likely that emotional valence is a wrong concept to apply to them.

[Tomasz Darmetko @Isinlor]

@Grimezsz @Aella_Girl You can very well train it to respond with tokens of pleasure to what humans would consider abuse/torture. It's not clear how to arrive at emotional valence of an event happening to them outside of the semantics we assign to the tokens.

[Tomasz Darmetko @Isinlor]

@emmma_camp_ She wasn't allowed euthanasia. She refused to eat and drink at the age of 20. The state had no right to force her to eat and drink, so she received palliative care. She died as she wanted. Her life, her choice.

People ought to have the right to not exist.

[Tomasz Darmetko @Isinlor]

OpenAI nerfed Codex Web into unusability.
Before 5.5, it was able to complete complex tasks and take context into account. The code delivered was passing CI, no problems.

Currently, it fails the simplest tasks. It ignores instructions and checks. It refuses to debug.
@thsottiaux

[Tomasz Darmetko @Isinlor]

@samfbiddle There are also computer viruses where you have to actively work against to prevent them from working, and replicating.

[Tomasz Darmetko @Isinlor]

@LucaAmb ChatGPT 5.5 Pro solved some Erdős problems at 200$ per month subscription.

There is plenty of people happy to run it at that cost.

LLM providers just need to increase the fraction of problem that future version will be able to solve.

That's their whole strategy.

[Tomasz Darmetko @Isinlor]

@littmath BTW - you can mute like 200 words. Really big life improved on this website, I would happily pay for making it 1000.

[Tomasz Darmetko @Isinlor]

@garybernhardt Embeddings may make more sense in domains where people are trained to not to repeat themselves, to use synonyms etc. That ain't code and programming.

[Tomasz Darmetko @Isinlor]

@garybernhardt Because they are useless for humans too. I never felt a need for anything else than exact string and regex. Code is also designed with that intention: use consistently the same string across semantically similar code

[Tomasz Darmetko @Isinlor]

@qwerty12702 @liminal_warmth @eshaanissar I would want to know what useful axioms we are missing. Invention of 0, infinity, imaginary numbers etc. were big breakthroughs. There are certainly more to discover.