Jhonathan P. R. dos Santos

Research Scientist at #Corteva (tweets are my own) | #Genetics, #PlantBreeding, #Analytics, #AI

[Jhonathan P. R. dos Santos @jhowpd]

@elonmusk Grok 4.1 thinking have been showing way worse performance then grok 4 fast in coding. Why ? 🤔

[Jhonathan P. R. dos Santos @jhowpd]

@elonmusk I think, probably have a completely different model to support grok 4 fast would be great. I noticed often he get stuck in a wrong solution direction. I switch between gpt-5 and grok 4 and solves the problem. Probably the best move is just by another competing, orthogonal product

[Jhonathan P. R. dos Santos @jhowpd]

@sama Reality is that Gemini 2.5 pro is just greatly better then o3. Is GPT5 that greatly better? What Google is shipping next. Is finally the fall of gpt series? Interesting to see what is coming.

[Jhonathan P. R. dos Santos @jhowpd]

@hot_girl_spring Code in C++ and bind.

[Jhonathan P. R. dos Santos @jhowpd]

@OX_DAO Probably I have no IQ, I can only code when I listen music.

[Jhonathan P. R. dos Santos @jhowpd]

@olson_dan You write as fast and memory safe code in C++. Just need to use modern C++, and still have a four decades possibilities if you need. Rust need many decades for that.

[Jhonathan P. R. dos Santos @jhowpd]

@benno_krojer @NandoDF There still a plenty of relevant content on biology on YouTube. I personally got started on Machine Learning a decade back watching Nando YouTube videos. Tons of Biology courses as well.

[Jhonathan P. R. dos Santos @jhowpd]

@ThePrimeagen Mix is the best. I particularly have huge technical steps by first fully reading a book (reading one now about GPUs), writing from scratch all book examples of different patterns, then migrate in existing projects for optimization. Best way to level up

[Jhonathan P. R. dos Santos @jhowpd]

@ThePrimeagen @SlutskyTom If you go just work with functional strongly typed components, preferably with predictable standard containers, custom types only when absolutely needed, and composed no more then one hierarchical level with standard containers. Then assert when state is created. You are good.

[Jhonathan P. R. dos Santos @jhowpd]

@bindureddy I think there will be room for all. Both Programar s and builders, like always. Programming it is just the imperative, safe way to do it.

[Jhonathan P. R. dos Santos @jhowpd]

Pro tip when playing with pybind11: if creating a custom optional type, make sure to use std::variant instead of std::optional, and create an empty struct to mimic the None from python. It will be seemless python, with has_value, value_or, set_value reset, and repr (basic) types.

[Jhonathan P. R. dos Santos @jhowpd]

why is so hard for openai create a context setup that once you asked only for giving code with no explanation, just remember this. How annoying is that.

[KanekoaTheGreat @KanekoaTheGreat]

.@ggreenwald highlights @elonmusk's opposition to Brazil's authoritarian censorship regime.

The censorship framework in Brazil mirrors tactics employed by the US government to suppress dissenting voices under the guise of combating disinformation.

"Any criticism of this judge that offends him will result in punishment, no trial, no due process, and that's what they've done now to Elon Musk."

"This censorship regime has no basis in law. They actually tried to pass a law several months ago that would democratically enact this censorship regime to formally define the powers and the responsibilities of big tech. But they could not get the votes in Congress to pass it."

Thus, Judge Alexandre de Moraes was given unilateral authority to dictate permissible online speech.

"When they issue these orders to these platforms, they say, we want this member of Congress banned, we want this journalist banned, we want this activist banned. They provide no due process. They provide no explanation. And these orders are never sent to the individuals who are being censored."

"There is no judicial process for contesting the order. If you do try to contest it, you go right to the judge who issued it, and obviously, you have no chance of being heard."

"It's not just censorship. It's censorship with no due process in the most Kafkaesque way; there's not even a legal framework for it."

"They're not even allowed to inform their users that they received a judicial order. Let alone disclose to the public or their users what the content of the order is. They just have to obey and ban the individuals named in the order, and if they don't do so within two hours of receipt, they will face daily fines of $20,000 for every account they fail to ban."

In January 2023, Greenwald obtained one of these judicial orders that revealed arbitrary bans on democratically elected congress members of the House and Senate, including 25-year-old politician Nikolas Ferreira, who garnered record-breaking support.

Ferreira not only secured a historic victory by amassing the largest vote total of any congressional candidate in Brazil's history but also garnered over a million votes in his state, surpassing all other candidates in the 2022 election.

However, despite his widespread support, the judge overseeing the censorship regime arbitrarily banned Ferreira from all online platforms, silenced him, and deprived him of his online presence without any explanation or opportunity for due process.

Senators, journalists, and activists who express dissenting opinions or criticize this censorship are similarly targeted for censorship by the same judge, highlighting the authoritarian nature of these actions.

"Needless to say, Google, YouTube, Facebook, Instagram, and WhatsApp, which is by far the most popular chat platform in Brazil, have always just meekly complied with every censorship order they got from Brazil, but Elon Musk decided he had enough."

[Jhonathan P. R. dos Santos @jhowpd]

Hack of the day: pass a list[int] from python as a binding with pybind11 and make it work in C++ without memory duplication and with int16 type, with seamless interaction from python. Lot of pain, but worked ;)

[Jhonathan P. R. dos Santos @jhowpd]

@NandoDF @Azure @OpenAI Pelo menos no GTC foi o que aparenta. Estou dedicando bastante tempo em CUDA C e transformers. Pra mim ta muito claro no futuro a infraestrutura mais rapida, vai gerar melhor inteligencia. Obrigado professor Nando, seu curso no YouTube aberto mudou minha vida 10 anos atras.

[Jhonathan P. R. dos Santos @jhowpd]

@sirbayes That's the beauty of it. Less is more.

[Jhonathan P. R. dos Santos @jhowpd]

kind of funny how GPT-4 holds back explanations about CUDA Graphs.

[Yann LeCun @ylecun]

Let me clear a *huge* misunderstanding here.
The generation of mostly realistic-looking videos from prompts *does not* indicate that a system understands the physical world.
Generation is very different from causal prediction from a world model.
The space of plausible videos is very large, and a video generation system merely needs to produce *one* sample to succeed.
The space of plausible continuations of a real video is *much* smaller, and generating a representative chunk of those is a much harder task, particularly when conditioned on an action.
Furthermore, generating those continuations would be not only expensive but totally pointless.
It's much more desirable to generate *abstract representations* of those continuations that eliminate details in the scene that are irrelevant to any action we might want to take.
That is the whole point behind the JEPA (Joint Embedding Predictive Architecture), which is *not generative* and makes predictions in representation space.
Our work on VICReg, I-JEPA, V-JEPA, and the works of others show that Joint Embedding architectures produce much better representations of visual inputs than generative architectures that reconstruct pixels (such as Variational AE, Masked AE, Denoising AE, etc).
When using the learned representations as inputs to a supervised head trained on downstream tasks (without fine tuning the backbone), Joint Embedding beats generative.

See the results table from the V-JEPA blog post or paper:
https://t.co/mfLvtvk8jj

[Jhonathan P. R. dos Santos @jhowpd]

Programming is akin to cleaning: Start somewhere, address the nearest issue, and progress to the next, aiming for utmost cleanliness. In coding, substitute 'dirt' with 'errors', you learned how to code.

[Jhonathan P. R. dos Santos @jhowpd]

@ylecun Just the amount of network data transfer they can safe alone, make them incredible useful, plus other plantora of utilities, like ML-ready representation, minimization of computing cost - associated with data transformation and so on.