gkrovso

content creator | my life | and my dreams

[gkrovso @bebraSOL]

Anthropic CEO:

"This is a super weapon. You should have to own a gun license to use it. Please don't release this."

In his interview, Dario Amodei explains why he is keeping the newest model, Mythos, hidden from the public.

Autonomously goes through all the links of the cyber kill chain. 271 new vulnerabilities found in Firefox. Enormous commercial losses from refusing a public release.

He admits that today's cyber safeguards can still be jailbroken too easily, and he is waiting for a time when the internet ecosystem becomes more secure.

The most responsible voice in AI, or a brilliant PR strategist who turned a product delay into the ultimate marketing?

[gkrovso @bebraSOL]

@xeophon @a1zhang @sethkarten @omouamoua @kevinjosethomas all the hard work you've done will definitely be recognized by the community

i believe in you

[gkrovso @bebraSOL]

Prime Intellect ( @PrimeIntellect ) just dropped Prime Agent, and it's a completely different way of building agents

instead of fixed tools and static prompts, the agent gets programmatic access to its own harness, context, history and sub-agents through a live IPython kernel

context becomes a variable it can manipulate directly, spawning sub-agents like function calls:

- Recursive Language Model harness, agent can rewrite its own scaffolding at runtime
- the only tool is a Python kernel, everything else runs as regular functions
- sub-agents run as persistent sessions and talk to each other directly
- self-improvement via /refine, agent edits its own harness with rollback available

the results are what's making noise:

- 95.5% Best@1 on ARC-AGI-3 with Opus 5, above human expert baseline (95.4%)
- 99.97% at Best@3
- built SEGA Genesis and Game Boy Color emulators from scratch in Rust
- production score 100k+ on Factorio Learning Environment

fully open-source, one-line install, and the gains show up on both open and closed models, even ones never trained for this harness

self-improving agents that rewrite their own architecture, or a research demo that won't hold up outside benchmarks?

[gkrovso @bebraSOL]

looks like Qwen 3.8 max is somewhere between Claude Opus 4.8 and Opus 5

or maybe Opus 4.6

at least based on their benchmarks

[gkrovso @bebraSOL]

@Polymarket wonder how many defendants this actually covers

[gkrovso @bebraSOL]

@notjazii shame it's blocked wondering if it's regional or a waitlist thing

[gkrovso @bebraSOL]

@_matisei blackrock and visa running validators is a wild shift

[gkrovso @bebraSOL]

@OpenMed_AI i believe in you

[gkrovso @bebraSOL]

@DanDr1s thanks Dan

[gkrovso @bebraSOL]

@svpino tool calling plus 1m context is a wild combo if it holds up

[gkrovso @bebraSOL]

@KyleDeWriter that's a huge jump from 9b in october 2025

[gkrovso @bebraSOL]

@AntLingAGI @ValsAI @huggingface @ModelScope2022 open weighting is a big step for ling 3.0 flash

[gkrovso @bebraSOL]

@testingcatalog new ai models would be a great announcement

[gkrovso @bebraSOL]

@KentonVarda interesting that it's built on cloudflare workers this time

[gkrovso @bebraSOL]

@0xkvro that 1 percent edge can add up over time

[gkrovso @bebraSOL]

@WatcherGuru what happens if they dont pass it now?

[gkrovso @bebraSOL]

@Oluwaphilemon1 how about comparing kimi k3 and gemini 3.6 flash?

[gkrovso @bebraSOL]

@0x_sakata i wonder how it handles memory management

[gkrovso @bebraSOL]

@daniel_mac8 sounds like claude didnt quite work out for you

[gkrovso @bebraSOL]

@DeItaone in house chip team is a bold move for anthropic

[gkrovso @bebraSOL]

@Kalshi wonder if other experts agree with his 1987 comparison

[gkrovso @bebraSOL]

@furiokao i wonder what he thinks the blind spot is