Adam V

Rethinking human-computer interaction, interested in crazy graph problems, evolutionary algorithms, hyperdimensional computing, and AI alignment.

[Adam V @AdamVcoding]

An ephemeral valence of
Mythical powers and Feeble man
Self-combusts into political precedent

An entity control/control entity problem
Expediting the acceleration of improvement
Neural coupling dei ex machina

[François Chollet @fchollet]

The end will begin when humanity turns away from humanity

[Adam V @AdamVcoding]

Just heard "I'll get a *clanker* to <inaudible> about it" on the London streets, and while I believe we should have more derogatory terms for AI, embodiment is not a requirement.

"Bithead!" is more appropriate.

[Adam V @AdamVcoding]

@tangled_zans There's the Scala one, and there used to be https://t.co/sXk1rmMI2W
Definitely a gap to be filled with a new FP meetup!

[Adam V @AdamVcoding]

@ID_AA_Carmack For neural computations where regularizers are ubiquitous, luckily, you're 4σ sure an activation will not land in the coarse (32 bins or less) >4 bf16 region

[Adam V @AdamVcoding]

@CPUMonkey AVX512_BMM is unironically the most exciting and powerful addition to AVX512, major congrats to AMD.
For those interested: https://t.co/4StMeijfJj

[Adam V @AdamVcoding]

Can you spot the transformer's heads, logits, and embeddings?

[Adam V @AdamVcoding]

Can you spot the transformer's heads, logits, and embeddings?

[Adam V @AdamVcoding]

@karpathy Great t-shirt design

[Adam V @AdamVcoding]

A small lab without the capacity to train a foundational model, can achieve SOTA with significant margin by using principled factored cognition: breaking up a task in a "well-typed" way not only increases efficiency, but also improves coherency on large tasks.

[George Morgan @vr4300]

I'm extremely proud to share that the @symbolica research team has achieved a monumental result in program synthesis. We have been able to reach SOTA on ARC-AGI-2 (85.28% @ $6.94/task) using @agenticasdk as a neurosymbolic program synthesis engine.

This engine is not ARC specific. It is 350 lines of highly generic code that can be readily adapted to any other task. This is what sets this result apart from other bespoke models or agent system designs that have historically performed well on ARC.

This result is a clear demonstration that the path forward to improve the reasoning capabilities of AI systems is by leveraging structure: types, composition, and program execution. Symbolic AI has laid dormant for decades but the field is on the precipice of making one of the greatest comebacks in history.

Blog: https://t.co/sFtjLfkMkf

Code: https://t.co/BZn74vbcXX

[Adam V @AdamVcoding]

@doganuraldesign Uncolor me unimpressed

[Adam V @AdamVcoding]

One of the last great achievements of humanity before AGI and AI-backed design

[SpaceX @SpaceX]

Splashdown confirmed! Congratulations to the entire SpaceX team on an exciting tenth flight test of Starship!

[Adam V @AdamVcoding]

@fchollet They are fun! And perhaps that's needed to complete them. There's more human bias in this set of environments and also more of an implied reward system (think curiosity, minimizing side effects, and temporal discount), and that may lead to some of the same problems humans have.

[Jared Isaacman @rookisaacman]

56 years ago, we achieved one of the most consequential moments in human history. We have waited patiently...it is time for the next giant leap.

[Adam V @AdamVcoding]

The most surreal hallucination I ever got from an image generation model. I do like Brussels waffles, but this is with a completely cleared ChatGPT personalization record. https://t.co/dEtm2yVSH7

[Adam V @AdamVcoding]

@insurrealist @prathyvsh @elzr What would those more general objects be?

[Adam V @AdamVcoding]

@danghica Analogously @u_map_prop ?

[Graham Hutton @haskellhutt]

@tritlo Monads are a simple algebraic structure with two operations and three laws. That’s it. No maths student ever complained that groups are too difficult to understand, and they have more things to remember. CS just need to grow up and stop demonising simple abstractions!

[Adam V @AdamVcoding]

@tangled_zans @_chenson__ Right. Not following Elon anymore helped for me. Although at a lower density than before, there's still many useful connections within tweet range here.