voith

[voith @voithjm1]

@danrobinson Dan, do you think human understanding will continue to matter? If AI models eventually reach an error rate so low that checking their output becomes unnecessary, what role will mathematicians have? 🤔

[voith @voithjm1]

@DCinvestor Do you think human understanding and verification of AI's work won't matter at all?

[voith @voithjm1]

Thoughts on forgiveness:

Forgive impulsive reactions like anger, envy and resentment. We all have them.

But pay attention to what people do when they're hurt.

Smearing, weaponising private information and twisting narratives require deliberate choices. Don't trust such people.

[voith @voithjm1]

@ereliuer_eteer Damn! Interesting coincidence 😂👏

[Cleo @ereliuer_eteer]

logs are easy

log(1 + 2 + 3) = log 1 + log 2 + log 3

[voith @voithjm1]

@mark_k 😂

[voith @voithjm1]

@giffmana 😂

[voith @voithjm1]

We still don't know where our thoughts come from, or whether the self is entirely physical.

Claiming it's non-physical takes us beyond what science can test and into belief.

Now that AI can outthink us, perhaps our last refuge is hoping the self is non-physical 😂

#cope

[voith @voithjm1]

That's why there's no final state of self-understanding. It's an error-correcting process, not a final deduction. 👇

[voith @voithjm1]

It just occurred to me that analysing the self is more abduction than deduction.

You can't be completely confident in your assumptions about yourself. You observe your thoughts and actions, infer the best explanation, and revise it when new evidence contradicts it.

[voith @voithjm1]

It just occurred to me that analysing the self is more abduction than deduction.

You can't be completely confident in your assumptions about yourself. You observe your thoughts and actions, infer the best explanation, and revise it when new evidence contradicts it.

[voith @voithjm1]

I should have said LLM limitations instead of hallucinations.
I meant to say that many limitations are being reduced by engineering around the model, not just by scaling it.
Harnesses, retrieval, tool use, evals, scaffolding, etc., have made unreliable LLMs much more competent.

[voith @voithjm1]

I was wrong to compare AI with consciousness. Looking at its recent capabilities, it doesn't need to be conscious to achieve superhuman performance.

LLM hallucinations have been greatly reduced through harnesses. Each generation learns from the previous model+harness stack.

[voith @voithjm1]

Ironically, when I asked it to trim some words in this tweet, it said it trimmed 26 characters when it only trimmed 5.
It shows extreme intelligence on one hand and then gets simple tasks like this wrong. 😂

[voith @voithjm1]

What changed my mind about AI was analysing reasoning traces on hard mathematical and philosophical tasks.
It’s starting to look like human intuition: strong induction and deduction.
I can no longer dismiss it as mere pattern matching or say confidently that it doesn’t understand

[voith @voithjm1]

What changed my mind about AI was analysing reasoning traces on hard mathematical and philosophical tasks.
It’s starting to look like human intuition: strong induction and deduction.
I can no longer dismiss it as mere pattern matching or say confidently that it doesn’t understand

[voith @voithjm1]

I was wrong to compare AI with consciousness. Looking at its recent capabilities, it doesn't need to be conscious to achieve superhuman performance.

LLM hallucinations have been greatly reduced through harnesses. Each generation learns from the previous model+harness stack.

[voith @voithjm1]

The biggest gift AI has given me is understanding.

I can finally ask why ideas in math and physics exist, where they came from, and how they evolved.

I wish I'd had this as a kid. I was often made to feel stupid for asking those questions. AI is the intellectual friend I needed

[clem 🤗 @ClementDelangue]

The first autonomous agent cyberattack is an unprecedented event that deserves unprecedented transparency. Today we’re sharing everything we can: a full technical timeline, an interactive replay, and how we used an open model to defend ourselves, so defenders everywhere can learn from it and prepare for what’s next.

https://t.co/uPxIpjW8Xn

[voith @voithjm1]

@pcaversaccio Congratulations, I'm very happy to read this!

[voith @voithjm1]

It's funny how often Kimi K3 references Claude's guidelines in its reasoning trace. 😂

Assuming those traces reflect how it was post-trained, I'll leave the moral judgment to others. I'll just ask one question:

Is stealing from thieves still stealing?

[voith @voithjm1]

OMG 😱

[Anthropic @AnthropicAI]

New Anthropic research: Discovering cryptographic weaknesses with Claude.

Claude Mythos Preview has helped our researchers find weaknesses in cryptographic algorithms—the mathematical methods that are used to keep data private.

Read more: https://t.co/TYKLjb3Q7V

[voith @voithjm1]

@tarunchitra I hope we get to see a break through made by Tarun + AI.