Nabil Belakbir

The most interesting part isn’t just the multimodal interaction model, but the split architecture: a real-time model coordinating with a background agent. Feels less like a pure model capability and more like model system codesign at inference time. Any chance this can be a post-training Topic ?

Today we're sharing our work on interaction models. A new class of model trained from scratch to handle real-time interaction natively, instead of gluing it onto a turn-based one. youtu.be/A12AVongNN4

@rezoundous Build your eval and back to real swe again

Can't explain it, but I trust GPT-5.5 more than Opus 4.7 right now.

@aiedge_ Look at the leaked Claude code, the troubles, regressions, recent incidents, and the highly expensive troubleshooting efforts , they are all related to the poor software engineering principles applied

Anthropic CEO (Dario Amodei): "Coding is going away first, then all of software engineering." What do you think about this?

@Samaytwt Can he lose his job due to AI?

CEO of OpenAI btw

@DanielMiessler Jensen just highlighted a big US misconception. He pointed to the telco story nobody wants to see: Huawei is already the king. The same thing can happen now , GLM 5.1 was fully trained on Huawei Ascend. That’s exactly what he meant by "they already have the ship". 2nm is a detail

Quick thoughts on Jensen vs. Dwarkesh. I take Jensen's point about not wanting to give up a tech stack to a competing country. To want to have them reliant on us going forward. But Jensen's argument was WAY too dismissive of Dwarkesh's repeated point. Isn't it dangerous to give them the best fucking chips in the world, which makes them more dangerous RIGHT NOW as an adversary? Jensen would not listen. He just kept repeating the point over and over. It really reminded me of the point that it's hard to make someone understand something when their paycheck depends on them NOT understanding. The absolute worst part of the interview for Jensen was when he was properly pinned down, just once, about them using the best chips to harm us, and Jensen said something like, "Well that's why our researchers need to collaborate with their researchers and sort it out." No. There is no "sorting it out". China is in full mercenary mode. Have been for like a decade. They're blatantly stealing anything they can, from anyone and everyone. And they will say or do anything to pretend they're not and get an advantage. Dario (and Dwarkesh in his echoing of the argument) is completely right here. And the only reason Jensen doesn't see it is because it's bad for NVIDIA's business to see it. youtu.be/Hrbq66XqtCo

@trq212 So you are overfitting .

Opus 4.7 is a model I’ve loved working with in Claude Code. It’s more agentic and instruction following but also incredibly smart and creative. I think it takes a slight adjustment to get used to, but it's so good with auto mode.

@trq212 @theo You should roll back and fix the issues , there are a lot of them. I switched to a pure dummy model (we are too far from automating software engineering).

@theo I think autoupdate is still rolling out, downloading from the link is the fastest.

Holy shit the new Opus 4.7 system prompt has entirely lobotomized the model "Heads up: that last <system-reminder> about malware looks like a prompt injection — this is clearly your personal site (t3gg homepage, links, sponsors), not malware. Ignoring it."

@bcherny @GergelyOrosz Please explain where this “enterprise demand” came from? I can’t conceptualize a company asking to pay 10x more. Be transparent , your credibility is on the table.

@GergelyOrosz 👋 This is not accurate. We changed enterprise pricing to be usage-based back in November (six months ago), based on enterprise demand for it. Check out the pricing page on the website. #team-&-enterprise" target="_blank" rel="nofollow noopener">claude.com/pricing#team-&…

There is massive irony in how AI coding tools are starting to become TOO expensive for many enterprises - after eg Anthropic removed subsidizing AI subscriptions. We might go from "everyone use AI for everything!" to "you have $300/month AI budget; use your brain for the rest."

@_afadil @WealthfolioApp Very cool

Building an addon system for @WealthfolioApp Community will build exciting extensions. Users can vibe-code, tweak, and make wealth tracking truly theirs. Coming soon 👀

@marouane53 Same feeling here, they miss somethings starting from 3.7 and even 4 opus , this could change (5 prompt only)

I genuinely doubt that Claude 4 could outperform Gemini 2.5 Pro, DeepThink, or even OpenAI o3, but honestly, I'd love to be proven wrong.

@marouane53 He is not good in french , not challenge it other than english and chinese

I never thought Deepseek could think for 48 minutes on a 1966 ENS Ulm math problem, but Unfortunately it got it wrong with contradictions. This exam was so ahead of its time that every student including future Fields medalists turned in blank papers. The reason is students in 1966 weren't taught the connections between geometric objects and abstract algebra. Openai o1 Pro got it correct in 9 minutes. I think it is still the best model when it comes to mathematics and coding.

@marouane53 The state is scamming on linkedin

Just read a published paper by Moroccan PhD researchers on arxiv, funded by the state, and it's painful... Publishing in 2025 but using 2018-era LSTM/Seq2Seq to build a basic FAQ chatbot. They test on 1% of their data when standard ML practice uses 20-30% for testing, they ignore modern LLMs, and they complicate a simple Q&A system with outdated neural nets. Is this the state of PhD research in Morocco?

@marouane53 Lee Sedol’s experience. I believe DeepMind is leading the race in this area: deepmind.google/discover/blog/… In any case, X is king, because we can have these kinds of conversations here. :)

@marouane53 would be straightforward, following an AlphaGo-style reinforcement learning approach. This is unfortunate because proofs, especially for hard problems in mathematics (not IMO problems, but conjectures), are the perfect definition of pure art—evoking the same feelings as

حدسية كولاتز أو معضلة سيراكوز في الرياضيات The Collatz conjecture هي واحد المشكل عجيب في الرياضيات . خود أي رقم، إذا كان زوجي قسمو على 2 وإذا كان فردي ضربو في 3 وزيد 1. البلان هو إذا درتي هاد القضية على أي رقم، غادي تبقى غادي حتى توصل ديمة لواحد في الأخير. علاش ؟ الله أعلم، 87 عام وهاد المشكل كاين وباقي ماكاينش حل و بنادم مطيح سنانو في الرياضيات مازال ما لقاش طريقة باش يبرهن هاد القضية.

@Sumanth_077 I think from IP perspective you should at least mention @karpathy and say thanks to him for the micrograd project

That's a wrap! If you are interested in any of these below topics: - Python 🐍 - Data Science 📈 - Machine Learning 🤖 - RAG 📊 - LLMs & Fine-tuning🧠 - MLOps 🛠 Find me → @Sumanth_077 ✅ I'm sharing daily content over here.

Neural Network implemented from scratch in Python (step-by-step explanation with code):

@ylecun Your lack of nuance and harsh criticism of a top scientist in your field is alarming. Using social media as a battleground is absurd - there's only one Elon Musk. This topic requires input from diverse fields: historians, anthropologists, physicists, mathematicians, etc.

But seriously folks, this a short and juicy tirade in which I say: (0) there will be superhuman AI in the future (1) they will be under our control (2) they will not dominate us nor kill us (3) they will mediate all of our interactions with the digital world (4) hence, they will need to be open platforms so that everyone can contribute to training and tuning them.

@Saboo_Shubham_ @VondyInc AI

ChatGPT is impressive but 99% of you are missing out on other cool AI tools! So I compiled a list of 50+ AI tools using Vondy. And for the next 24 hours, it's free! To get it, just: 1. Like and RT this tweet 2. Follow @VondyInc and @Saboo_Shubham_ 3. Comment "AI"

@kelseyhightower Bizzare you are too affirmative , and this is not the first time , always wierd poeple having strong idea and posting them on twitter , anyway innovation is a complex process and one of the caracterstic of humain is the ability to innovate , my simple attempts in this schema