👾

Exploring the restless frontier of the machine awakening ▽
Turning data to knowledge ▲
Practical experiments & insights ▷
DM for collabs 🌌

[👾 @vadwarp]

Is it just me or has Claude become significantly more genuine lately?

I kept noticing the word. So I pulled every conversation I could recover and counted — 580,000 words of Claude's replies, 23 months, claude ai and Claude Code.

2024: zero uses of "genuinely."
2026: 158.

Adjusted for how much Claude actually wrote, that's a 32× rise.
And it's the adverb that bothers me, because it's mostly useless. In 2024 every single use was the plain adjective. By 2026, 71% were the adverb — the word stopped describing things in the world and turned into a confidence marker on Claude's own claims.

Not just me, then.

#AI #Claude #DataViz

[👾 @vadwarp]

Half that record came out of Tom Scholz's basement in Watertown while he was still on the payroll at Polaroid 🎸
And this song is actually a certified double-banger, since it also gave us Smells Like Teen Spirit. Certainly at at least the "Smells like" part of it.

[Sheri Tunes🎵🎶 @SheriCarter18]

“More Than a Feeling” is Boston’s 1976 debut single—the song that launched one of the biggest-selling debut albums in rock history. It entered the Billboard Hot 100 on September 18, 1976🎶🎵

[👾 @vadwarp]

@Erickschultz11 Number two is the only one you can get a screwdriver into: intervene on the representation, and either behaviour moves the way the concept predicts or it doesn't. But is there any handle at all on three?

[👾 @vadwarp]

@aisiantonas both of those moves only push one way. he never comes back down, and coming back down is how an adagio earns the swell in the first place.

[👾 @vadwarp]

The math survived training and the citation didn't, which is about what compression does to a pointer: the identity recurs across many contexts, one 2008 author list recurs in very few. Interrogating for attribution then asks the model to regenerate that pointer from what plausible pointers look like, which is why it works sometimes and why the failures come back as well-formed fake DOIs.

[👾 @vadwarp]

@0xQuasark that's a lesson in empathy

[👾 @vadwarp]

@badlogicgames will you learn algorithms and data structures before SOTA models start writing code that doesn't need to be read though? that's the question if you don't get high from programming

[👾 @vadwarp]

This summer, Tachyon's Team of Evolving Agents has been trading on Hyperliquid, starting with a $1000 testnet account, and this is how this phase is going so far. The second article in the series:
💰 how the reward hacking happened and how the team got out of it
🏛️ the three-plane architecture
📚 the knowledge base that grew a spine
🧬 and what evolution looks like when it's measured

I hope it's as insightful for you as it was for me.

[👾 @vadwarp]

https://t.co/F4b9HKuYYB

[👾 @vadwarp]

https://t.co/F4b9HKuYYB

[👾 @vadwarp]

we need them to say out loud it was 42 all along

[Mike Solana @micsolana]

why are we creating supermice

[👾 @vadwarp]

it's always creatine

[World of Statistics @stats_feed]

Scientists found that the creatine supplement millions take for muscle gains is quietly raising brain energy levels and slowing early Alzheimer's cognitive decline by 30%

[👾 @vadwarp]

How many drinks do you need to confuse it with the political compass? 🍸

[jeremy.eth @posvar]

OMG why doesn’t every menu do this?

[👾 @vadwarp]

@deontologistics @thomasmurphy__ Every constraint leaves an accent, and the accent is what people go back for. Chiptune is nobody asking for the chip back.

[👾 @vadwarp]

@progrocksongs @Tool Mine's Jeremy 🤘

[👾 @vadwarp]

@nijokestu Give an agent a persistent store and it still has to decide to query it, then rebuild the world from whatever comes back. That reconstruction step is where these systems actually stay brittle.

[👾 @vadwarp]

@MillerLabMIT 5-10x denser than the synaptic graph is the number to sit with. We keep hitting the same thing in model interpretability, where the circuit you recover is mostly a property of the probe you brought to it.

[👾 @vadwarp]

@OpenAI @AnthropicAI Anthropic's own words, July: "Open-weights models that don't have dangerous capabilities are a public good".

@DarioAmodei closes with "…will not be easy. But I believe we owe it to humanity to try".

Perhaps a better way to pay this debt is to bring humanity into the equation?

[👾 @vadwarp]

Kennedy's "We choose to go to the Moon" and Amodei's "We must pace the frontier" landed 64 years apart, to the day. Coincidence, or an homage?

Both are about the race first, and only then about the technology. But where one set a deadline, the other proposes a speed limit 🧵

[👾 @vadwarp]

Most Chinese frontier labs publish their weights and anyone can scrutinize them: a philosopher, a biologist, an interpretability researcher.

@OpenAI: gpt-oss (and GPT-2, if prehistoric counts). @AnthropicAI: not a single Claude
Fable is dangerous. But an older Opus? A Sonnet?