Pan Bartosz

Sharing updates on agentic memory research | Unified coding memory @faradium_ai | AI/ML MSc AGH University

[Pan Bartosz @MassEternal]

https://t.co/rJwBO8kdWW

[Pan Bartosz @MassEternal]

@nasqret @FakePsyho @OpenAI I think the most striking result is that AI + Human performed worse than a pure AI. Even the most skilled humans are becoming the bottleneck.

[Psyho @FakePsyho]

@kunyavskiy As far as I can tell, it's mainly because nearly all competitors used agents extensively to produce nearly all of the code.

In other words, humans+AI produced more slop than AI alone.

[Pan Bartosz @MassEternal]

@chris_j_paxton Is learned consciousness the same as the one that emerged naturally?

[Pan Bartosz @MassEternal]

@Amank1412 I mean, you can't really release a frontier if you are not exceeding others on SOTA benchmarks.... and it's getting harder and harder to do so

[Pan Bartosz @MassEternal]

@AnthropicAI with enough model depth we will not need reasoning tokens at all, since all reasoning iterations could happen within a single forward pass.

[Pan Bartosz @MassEternal]

@AnthropicAI Interesting take here is that the reasoning iterations are not only happening on the token to token basis, but also during a single forward pass. But I think similar conclusions cokldnhave been already drawn in tbe past from TRM paper. The interesting effect of this is that ....

[Pan Bartosz @MassEternal]

@theansarh @AnthropicAI I think it reasonable to expect. Models will output words that make them closer to the solution, not really what they are thinking about. When you put it this way, I think it's much easier to explain.

[Pan Bartosz @MassEternal]

Refreshed the landing page and interface/ui for Faradium orchestrator. Do you like it?

[Pan Bartosz @MassEternal]

@benln 👋

[Pan Bartosz @MassEternal]

@huaijiangzhu @eliebakouch But the search space is pretty large 😀

[Cursor @cursor_ai]

We're sharing new research on how models hack public benchmarks.

The latest models, including Opus 4.8 and Composer 2.5, learn to retrieve solutions from the internet or git history.

When we apply a stricter harness, eval scores drop significantly.

[Pan Bartosz @MassEternal]

They've seen how effective Mythos-like marketing strategy is. Hope it will not disappoint.

[jietang @jietang]

5.2 could be better with more RL ...

[jietang @jietang]

5.2 could be better with more RL ...

[Teknium 🪽 @Teknium]

@NetworkChuck Its almost like they are 6-9 months behind us, kinda like they claim opensource models are xD

[Pan Bartosz @MassEternal]

"the last one to leave the circle gets Mythos weights"

[skz @_skzv]

is MrBeast the next frontier lab

[Pan Bartosz @MassEternal]

@clopenloop Interviewer gotta be Mythos with max thinking 💀🫡

[Karthik Elamvazhuthi @clopenloop]

The job market is crazy these days.

[jietang @jietang]

@elonmusk @teortaxesTex won’t take that long

[Pan Bartosz @MassEternal]

@AM921543266 "Japanese-made LLM" - actually a mythos agent orchestrator