Amit Levy

founder

[Amit Levy @amitlevy64]

@distributionat I did

[Amit Levy @amitlevy64]

@mahdi_tcs_ Interesting that no one updated the Wikipedia until I did

[Amit Levy @amitlevy64]

@tshevl @Polymarket @Kalshi @_Mantic_AI METR is the wrong analogy, the markets themselves are rapidly becoming more accurate because these very systems are trading on them. Mantic is approximately level with FutureSearch which is a public API, as long as that's true it can't reach $1M volume accuracy.

[Amit Levy @amitlevy64]

@tszzl Not for long

[Amit Levy @amitlevy64]

@patrykperduta Hey, I want to publish a short rhythm game with this as the audio track, of course linking to your YT and crediting you. Is that OK with you?

[Amit Levy @amitlevy64]

Astra doesn't beat Fable 5.1 on code or forecasting (the things actually important for RSI), the reason the benchmarks feel that way is that any release includes only the winning benchmarks

[Pratham @Prathkum]

I have a genuine question.

Anthropic has probably been working on Fable 5.1 for a very long time. OpenAI has probably been working on GPT-6 Astra for just as long. These labs are almost certainly building in parallel, not reacting to each other in real time of course.

So why does Fable 5.1 launch just a couple of days before Astra, and then Astra beats it on almost every benchmark by a good margin?

And this isn't a one-off.

Every single launch follows the same pattern: when Anthropic ships a new model, it beats all of OpenAI's previous models.
When OpenAI ships a new model, it beats all of Anthropic's previous models. Now they are actually beating what the other just launched days earlier.

Benchmarks are completely vanity metrics.

[Amit Levy @amitlevy64]

@TheZvi It's very good on my internal forecasting benchmark, much better than previous SOTA (Astra is weak)

[Amit Levy @amitlevy64]

pretty cool!
https://t.co/24PZWc4kU4

[Amit Levy @amitlevy64]

@scaling01 this is a different sort of sacrifice vs what you seem to be describing in your quote, not sacrificing short term gain for long term gain, but sacrificing your own utility for another's

[HomeDAO 🏴‍☠️ @homeDAO]

Whom shall I send? And who will go for us?

homeDAO cohort 3 - applications now open

[Amit Levy @amitlevy64]

@timfduffy I doubt they are doing this in practice, they are scaring people away from using Claude for frontier work by simply suggesting silent sabotage as a possibility

[Amit Levy @amitlevy64]

@LatosLouis @dwarkesh_sp You can implement backpropogation as a local operation, the gradient at each node depends only on the gradients of the node it is connected to in the next layer (i.e. has non zero weight toward)

[Amit Levy @amitlevy64]

@LatosLouis @dwarkesh_sp It is absolutely local

[Amit Levy @amitlevy64]

@AmitLeViAI Bro someone just messaged me thinking I am you 😭 too many Amit Levy

Also one good paper will get you more citations than 10 bad ones

[Amit Levy @amitlevy64]

for me that's a downgrade

[Amit Levy @amitlevy64]

fuck, opus 4.6 alignment is bad. it keeps lying

[Amit Levy @amitlevy64]

What did they do to Gemini

[Amit Levy @amitlevy64]

@jamortenson @emollick You need the Claude chrome extension, once it's downloaded ask Claude to set up the connection with it

[Amit Levy @amitlevy64]

@chris_j_paxton Anthropic has a strong edge in coding right now

[Amit Levy @amitlevy64]

@javilopen @nateliason This is a meme, it's just a factorio screenshot