Eye01

Half machine, half human. AI Engineer/Artist.
Pictures via custom AI algorithms/artwork.
Exploring new ways to make art since 2017

[Eye01 @Eye01_]

@reve For info, your website crashes as it uses baseline 2024 accessory javascript functions, cutting out from some prior env contexts.
Uncaught TypeError: Promise.withResolvers is not a function
at new <anonymous> (reveapi-CMLtN4ts.js:3:22132)
at reveapi-CMLtN4ts.js:3:21769

[Eye01 @Eye01_]

@MacLesggy Nous avions déjà réfléchi sur ce système il y a 8 ans avec un autre ingénieur. La puissance du laser nécessaire risque de brûler la rétine dans de nombreux cas (erreurs, humain/animal dans l'axe, r��fraction/réflexion...). Même en triangulant/ répartissant la puissance sur n srcs.

[Eye01 @Eye01_]

@arcprize @jeremyberman @_eric_pang_ Congrats for the nice result ! I looked at your code @jeremyberman but if I may say so, I wouldn't call this "evolutionary" as it lacks basic key concepts from that field. It's more a "self-tuning trial-and-error prompt engineering" on different LLMs to get the best of them.

[Eye01 @Eye01_]

@elonmuskTN 39

[Eye01 @Eye01_]

@VictorTaelin @UnderwaterBepis They are not only CUDA binders... And you'll get a much faster result, which matters for big projects, if you keep a transparent intermediary step generating CUDA.
Also, just a notice : The legend about not using x% of our brain, is totally wrong.

[Eye01 @Eye01_]

Atomic Splash (HDversion) with music on IG
(Post from 2023- Nov-5)
https://t.co/iEasxXzjZm

[Eye01 @Eye01_]

@paultrillo @OpenAI @realwashedout Consistency problem... The couple has not the same faces at each change of shot. Couldn't you train a Lora instead?

[Eye01 @Eye01_]

@ylecun Si le référentiel est héliocentrique, il faut tenir compte
-de la rotation de la terre (axe incliné de 11.5°)
-et de son orbite autour du soleil
La vitesse orbitale étant de 29.78 km/s, les autres vitesses sont négligeables.Le point P devient spatial et hors de portée.Réponse 4)

[Eye01 @Eye01_]

@ylecun La réponse dépend du référentiel.
Si P est dans un référentiel géocentrique alors la réponse est 3)
La longueur parcourue D est la circonférence d'un cercle de rayon R'=R.sin(k/R) avec k la longueur de l'arc et R le rayon de la terre. donc D=2Pi.6378.sin(1/6378) < 2Pi km

[Eye01 @Eye01_]

4K version
https://t.co/REjazTDhzk

[Eye01 @Eye01_]

@RemiLeblond @GoogleDeepMind Anyway congrats to all team, and unpatient to see what it will give with fine-tuned Gemini Ultra on both scoring model and policy models !

[Eye01 @Eye01_]

@RemiLeblond @GoogleDeepMind It could be also interesting to learn with some extra parameters to select the code generation, as the challenge is not only to find a good result but how we obtain it. Ex : efficiency(speed/size), maintenance ("readability", i.e. declared local variables...), portability...

[Eye01 @Eye01_]

@RemiLeblond @GoogleDeepMind Nice ! How many attempts for this example ? (AlphaCode 2 Tech Report mentions 10 submissions in the pipeline, but how to rate this ? 5/10 where with good score ? or 1/10 ?)

[Eye01 @Eye01_]

Résultats très impressionnants pour Gemini Ultra ! Congrats to @GoogleDeepMind @GoogleResearch
@OriolVinyalsML @JeffDean

[Jeff Dean @JeffDean]

On a wide variety of text, math, coding and reasoning benchmarks, including MMLU, GSM8K, MATH, Big-Bench Hard, HumanEval, Natural2Code, DROP, and WMT23, Gemini outperforms all other models and improves on the state-of-the-art. On MMLU, Gemini Ultra is the first model to achieve human-expert performance on MMLU across 57 subjects with a score above 90%.

[Jeff Dean @JeffDean]

Gemini models build on top of Transformer decoders that are enhanced with improvements in architecture and model optimization to enable stable training at scale and optimized inference on Google’s Tensor Processing Units. They are trained to support 32k context length, employing efficient attention mechanisms (for e.g. multi-query attention (Shazeer, 2019)).

Gemini comes in a variety of sizes, including two flavors of Nano (designed for on-device usage in memory-constrained settings like phones), Pro (for enhanced performance and deployability at scale in datacenters), and Ultra (a for highly-complex tasks). Even the smaller models are pretty great!

[Eye01 @Eye01_]

@ylecun Very good, except for the watermarking, which is an intellectual regression stamp. A certain kind of political diktat (censorship, control, tracking, sanctions...) should not be automatically included in a science tool able to make creation. Pictures or sounds should remain free.

[Eye01 @Eye01_]

4K link on Youtube :
https://t.co/O7VBoBTaVx

[Eye01 @Eye01_]

@ylecun Personally, I think that any attempts to reason about the future actions of a possible superintelligence are completely futile and subjective. If it surpasses us, it will have mechanisms far too complex for us to understand. Can ants understand Einstein's theory of relativity ?

[Eye01 @Eye01_]

@ylecun @fltmu Open source including open data / model access ?

[Eye01 @Eye01_]

@bindureddy Great but is it possible to stop talking about AGI ?🙏 Replace AGI word by Terminator T800, or Skynet, or Darth vador, or any villains in any movie, that will give a good idea about how stupid looks this overtrending use of AGI term in all (interesting) talks about real AI