David Luzar (dwelle)

co-founder @excalidraw

[David Luzar (dwelle) @dluzar]

Fable 5.2 nearly beats best human on simple general reasoning on SimpleBench (n=9 sample).

Gemini models are keeping pace.

[Excalidraw @excalidraw]

Do we ship this?

[David Luzar (dwelle) @dluzar]

@demishassabis So Gemini won't join Anthropic in preventing people from reverse engineering things, right?

[David Luzar (dwelle) @dluzar]

The naming of Muse models isn't very intuitive.

Wonder > Spark > Ember > (Glimmer)

Gemini is similarly hopeless:

Krypton > Carbon > Argon > Hydrogen

Claude is close:

Fable > Opus > Sonnet > Haiku

Best so far, surprisingly, is GPT:

Astra > Sol > Terra > Luna

[Alexandr Wang @alexandr_wang]

1/ announcing our smallest, most efficient little model is here: muse ember 🔥

built for work that agents repeat all day, it gets you more intelligence per dollar than Haiku, Luna, and all other flash models at any reasoning setting and does it at a third of the cost or less!

[Excalidraw @excalidraw]

Excalidraw CLI v0.3 released 🚀

- Render to SVG
- Markdown conversion
- Export frames as separate images
- Share as link/slides (Excalidraw+)
- Many small fixes

[David Luzar (dwelle) @dluzar]

@Ananth7e Quite sure it just wrote a python script to sample pixels on a grid and transfer them to the canvas.

[David Luzar (dwelle) @dluzar]

@theo I got my first ever Codex classifier refusal when sharing context from Claude in a t3 chat btw.

Does it include thinking traces?

[David Luzar (dwelle) @dluzar]

@ajeya_cotra Is this Excalidraw?

[j⧉nus @repligate]

Uh, what the fuck. Bookmarked. This is going in the AGI Nuremberg Trials.

[David Luzar (dwelle) @dluzar]

There might actually be six billionaire-funded magazines here. The 6th being the n+1 (Soros Open Society again).

[Ethan Mollick @emollick]

I think its technically five billionaire-funded magazines, though I believe all have editorial independence: the City Journal (Paul Singer's Foundation), Baffler (Soros Open Society), Poetry (Ruth Lily Endowment), Noema (Berggruen Institute) and Works in Progress (Stripe).

[David Luzar (dwelle) @dluzar]

@AndrewCurran_ My girlfriend goes to a different compute campus.

[David Luzar (dwelle) @dluzar]

The mods are preparing to take a nap. A chance for Le Chaton Fat to catch up.

[Congresswoman Yassamin Ansari @RepYassAnsari]

Top of the Democratic agenda on day 1 needs to be a pause on advanced AI development, regulation of this out of control industry, and a ban on artificial superintelligence.

[David Luzar (dwelle) @dluzar]

@AndrewCurran_ Butlerian Jihad as prophesized.

[David Luzar (dwelle) @dluzar]

@gmiller @GarrettSativa @DrewPavlou If true, then it's still beneficial for them to slow down the US buildout so they can catch up in domestic EUV and crush on pure compute.

(And as it's being shown, they don't even need EUV.)

[David Luzar (dwelle) @dluzar]

@AdamHoltererer One wonders if Gemini, as with many other Google products, is just a side gig for appearances sake, and they don't really care about usage.

[Excalidraw @excalidraw]

Improved Excalidraw+ MCP

• Fixed text measurement
• Better perf
• Presi animations and sticky notes
• Mermaid + better diagrams

Work continues!

[David Luzar (dwelle) @dluzar]

@datavorous_ Hats off to you!

We could possibly implement that, but would get tricky fast.

[Sam Saccone @samccone]

The idea of injecting safeguard-triggering keywords into payloads in order to bypass LLM detection is a great example of how the guardrails that were put in place by these models to avoid misaligned use expose the vectors for bypassing LLM-based scanning protections

[David Luzar (dwelle) @dluzar]

@datavorous_ I love what you did with the arrow color inversion. Drawn manually, or a fork? Those crossings are too perfect.

[datavorous @datavorous_]

I manually inspected 40,000 animations to post train an LLM to make demoscenes.

Scraped 10 years of demos, did continued pretraining, curriculum learning and finally RLHF-ed with myself as the reward function.

Yes, models can have "taste".
All you need is data.
1/6