Software developer working on startups 💻 Triathlete 🏊🚲🏃
London, England
Joined November 2009
- Tweets870
- Following264
- Followers174
- Likes190
Luis Da Silva retweeted
Today we're introducing TRIBE v2 (Trimodal Brain Encoder), a foundation model trained to predict how the human brain responds to almost any sight or sound.
Building on our Algonauts 2025 award-winning architecture, TRIBE v2 draws on 500+ hours of fMRI recordings from 700+ people to create a digital twin of neural activity and enable zero-shot predictions for new subjects, languages, and tasks.
Try the demo and learn more here: go.meta.me/tribe2
Luis Da Silva retweeted
The invisible Glass experiment
Scientists once placed a transparent glass barrier inside an aquarium.
On one side was a fierce pike, and on the other side were several smaller fish swimming freely.
When the hungry pike saw the smaller fish, it immediately rushed forward to attack.
Bang. It slammed straight into the glass and bounced back.
Confused, the pike kept trying again and again, but every attempt ended the same way.
The repeated collisions injured its head and knocked off some of its scales. Eventually, the pike became frightened and retreated to a corner of the tank.
After some time, the scientists quietly removed the glass barrier. The smaller fish now swam freely throughout the aquarium, even brushing against the pike’s mouth.
But the pike never tried to eat them again.
Even though it was hungry, it refused to attack. In its mind, the invisible wall was still there.
A few days later, the pike reportedly died of starvation, surrounded by food. This phenomenon is often referred to as the Pike Effect or Pike Syndrome.
It’s often used as a metaphor for how repeated failure can create invisible limits in the mind.
I'm on the road towards completing my first ironman in 2027. Im thinking about sharing the journey here. Is that a good idea? 👀
Luis Da Silva retweeted
Elon Musk came up with a pretty incredible idea during the Q3 Earnings Call, that no one is really talking about.
His words: “Actually, one of the things I thought, if we've got all these cars that maybe are bored, while they're sort of, if they are bored, we could actually have a giant distributed inference fleet and say, if they're not actively driving, let's just have a giant distributed inference fleet.
At some point, if you've got tens of millions of cars in the fleet, or maybe at some point 100 million cars in the fleet, and let's say they had at that point, I don't know, a kilowatt of inference capability, of high-performance inference capability, that's 100 gigawatts of inference distributed with power and cooling taken, with cooling and power conversion taken care of. That seems like a pretty significant asset.”
So basically, each car has ~1 kilowatt of high-performance AI inference capability, Tesla wouldn’t need to build giant data centers — the fleet is the data center.
Tesla could turn their entire fleet into a giant distributed inference network, spread across the world, powered by the batteries and AI in the car already.
Mind blown.
Luis Da Silva retweeted
LLMs are injective and invertible.
In our new paper, we show that different prompts always map to different embeddings, and this property can be used to recover input tokens from individual embeddings in latent space.
(1/6)
Luis Da Silva retweeted
Finally, Python 3.14 lets you disable GIL!
It's a big deal because earlier, even if you wrote multi-threaded code, Python could only run one thread at a time, giving no performance benefit.
But now, Python can run your multi-threaded code in parallel.
And uv fully supports it!
Luis Da Silva retweeted
The end of Chain-of-Thought?
This new reasoning method cuts inference time by 80% while keeping accuracy above 90%.
Chain-of-Draft (CoD) is a new prompting strategy that replaces Chain-of-Thought outputs with short, dense drafts for each reasoning step.
Achieves 91% accuracy on GSM8k with ~80% fewer tokens than CoT
Luis Da Silva retweeted
OpenAI Strawberry (o1) is out! We are finally seeing the paradigm of inference-time scaling popularized and deployed in production. As Sutton said in the Bitter Lesson, there're only 2 techniques that scale indefinitely with compute: learning & search. It's time to shift focus to the latter.
1. You don't need a huge model to perform reasoning. Lots of parameters are dedicated to memorizing facts, in order to perform well in benchmarks like trivia QA. It is possible to factor out reasoning from knowledge, i.e. a small "reasoning core" that knows how to call tools like browser and code verifier. Pre-training compute may be decreased.
2. A huge amount of compute is shifted to serving inference instead of pre/post-training. LLMs are text-based simulators. By rolling out many possible strategies and scenarios in the simulator, the model will eventually converge to good solutions. The process is a well-studied problem like AlphaGo's monte carlo tree search (MCTS).
3. OpenAI must have figured out the inference scaling law a long time ago, which academia is just recently discovering. Two papers came out on Arxiv a week apart last month:
- Large Language Monkeys: Scaling Inference Compute with Repeated Sampling. Brown et al. finds that DeepSeek-Coder increases from 15.9% with one sample to 56% with 250 samples on SWE-Bench, beating Sonnet-3.5.
- Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters. Snell et al. finds that PaLM 2-S beats a 14x larger model on MATH with test-time search.
4. Productionizing o1 is much harder than nailing the academic benchmarks. For reasoning problems in the wild, how to decide when to stop searching? What's the reward function? Success criterion? When to call tools like code interpreter in the loop? How to factor in the compute cost of those CPU processes? Their research post didn't share much.
5. Strawberry easily becomes a data flywheel. If the answer is correct, the entire search trace becomes a mini dataset of training examples, which contain both positive and negative rewards.
This in turn improves the reasoning core for future versions of GPT, similar to how AlphaGo’s value network — used to evaluate quality of each board position — improves as MCTS generates more and more refined training data.
Luis Da Silva retweeted
It is time for the people of Venezuela to have the chance for a better future.
Support Maria Corina!
Tuve una grata conversación hoy con @JMilei, Presidente de Argentina🇦🇷 a quien agradecí por el apoyo y compromiso de su gobierno y del equipo político de la Cancillería durante estos meses tan difíciles para los venezolanos y para nuestra gente en los comandos de campaña.
Le reiteré mi confianza en el inquebrantable espíritu libertario de los venezolanos y nuestra convicción de que este 28J será una gesta cívica histórica.
El presidente Milei me reafirmó su apoyo a nuestra causa, a los valores democráticos y a la libertad.
Gracias, estimado Presidente.
Last week demanded my team and I to make the disheartening decision of closing our startup, Lagrange Systems....
These are some lessons from my story of how starting from the end led us to failure.
linkedin.com/pulse/starting-…
Life is all about giving.
Help as much and as frequently as you can.
Leverage your ways of helping as many people as possible.
Don't expect anything in return, just continue to help.
Help will eventually find its way back to you.
In many aspects of life, small differences in performance can lead to disproportionately large differences in outcomes.
E.g. Better products with better acknowledgement get disproportionately more customers.
While linearity is easy, superlinearity is way more common.
Have you heard of the Horsey Horseless?
It is a patent filed in 1899 for a wooden horse head that can be attached to the front of a car.
It was meant to make cars less spooky for horses.
A nice example of people underestimating the impact technology has on the future.
Luis Da Silva retweeted
NEWS: Elon Musk announced tonight that the first human implanted with @Neuralink’s brain chip has made a full recovery.
The patient is able to control a mouse using only their thoughts. Incredible achievement!
I enjoyed reading this paper on "Self-Discover", a new reasoning framework for LLMs (e.g. ChatGPT) just published by DeepMind.
Their results suggest this framework improves results by as much as 32% compared to Chain of Thought (CoT) 😯
I break it down a bit 🧵
It tends to be even better to use Plan-and-Solve, a two-step process where we ask the LLM to:
1. Create a plan on how to get the answer
2. Generate the answer by executing the plan.
The result is the LLM first describes the steps it should take and then takes them one by one.
Deepmind proposed "Self-Discover", where they prompt the LLM for a "reasoning structure" to gather the main knowledge and skills the LLM may use to solve the task to later connect them in the final answer.
To all interested, here's the link!
arxiv.org/abs/2402.03620
