Hamid R. Darabi

Do you ever feel tense when "vibe coding"? I used to experience that too. As a senior engineer and manager, weekends are a time for fun and learning through writing small pieces of code using GitHub Copilot. However, there was always a sense of incompleteness, ...

@nalidoust This guy has been a mouthpiece for the regime for the past few years, most likely taking money from them. No one takes him seriously.

The Islamic Republic regime has just arrested Parnian Khodabakhshi, a 20-year-old student at Aryamehr University, after security forces raided her home. Regime-linked channels announced her arrest alongside an image of a glass soda bottle, which is a threat of sexual assault against detainees in Iran. Pro-regime channels have claimed she is arrested solely because she carried the Sun & Lion flag during a peaceful university protest. Weaponizing the fear of rape to intimidate young women and silence student activism is cruelty by design. The world must not look away. Students demanding peaceful change should not face home raids, imprisonment, and sexual threats. #ParnianKhodabakhshi

That’s a very thoughtful question, Sahar. I think step 3 highlights that AI is inherently not very good at generating alternative approaches. This is particularly limiting when you need to make nuanced decisions because you are operating within a legacy infrastructure that imposes certain constraints. I hope that helped.

@_hdarabi Great thread. Treating Copilot like a junior dev + structured workflow really turns chaotic vibe coding into calm, productive sessions. Love the code in peace vibe. Which step in your 7 step process gave you the biggest quality boost?

Do you ever feel tense when "vibe coding"? I used to experience that too. As a senior engineer and manager, weekends are a time for fun and learning through writing small pieces of code using GitHub Copilot. However, there was always a sense of incompleteness, ...

5. Write tests with assistance, emphasizing key paths and failure modes. 6. Conduct detailed code reviews on these tests. 7. Provide clear next steps and continue until the feature is complete. With this structured approach, vibe coding became very productive. Code in peace!

my workflow transformed: 1. Define project context in an instructions.md file. 2. Translate business logic into high-level coding logic. 3. Explore alternative designs using the model. 4. Break down the problem into logical steps.

@Hesamation Sponsored post?

she actually summarized everything you must know from the “AI Engineering” book in 76 minutes. if you don’t got the time to read the book, you need to watch this. foundational models, evaluation, prompt engineering, RAG, memory, fine-tuning and many more. great starting point.

@NoahEpstein_ Show me the apps built with it.

cursor just made every $200/hour dev shop look like a clown dropped composer 2.0 yesterday with agentic browser built in what used to take 8 devs and 3 weeks now takes 8 AI agents running parallel in 30 seconds and they TEST THEIR OWN CODE in a native browser while coding bootcamps are charging $15K to teach you react, cursor's teaching AI to: → write code 4x faster than gpt-5 → run 8 versions simultaneously to pick the best one → test in chrome devtools without leaving the IDE → iterate on bugs until they're actually fixed → plan with one model, build with another the entire "hire a dev team" industry is sweating some startup just replaced 3 junior devs ($450K/year) with cursor pro ($240/year) that's a 99.9% cost reduction for better output the intelligence gap between "we staffed up our eng team" and "we deployed cursor 2.0" is getting stupid most companies still paying $150K/year for developers to do what this does for $20/month chatgpt atlas? cooked dia and comet? obsolete traditional dev shops? praying you don't find out about this comment "COMPOSER" and i'll send the full breakdown of how to replace half your dev costs with 8 parallel agents your competition is still hiring. time to bury them.

@james_mcwalter @GergelyOrosz Hi James, do you have any insights on what percentage is AI-generated?

@GergelyOrosz +23,000 applications in the last 30 days for 8 open roles, in person NYC.

I'm researching what is happening in the tech jobs market. If you're a hiring manager recruiting, or an experienced eng on the job market, would love to hear what you see. DMs open. (Feels like a weird market. I suspect data on eg jobs doesn't reflect what's on the ground)

Ever wonder how to make GenAI smarter? The short answer: make it larger. But how, and why does that translate into billions of dollars spent on AI infrastructure? I break it down in my new article. If you’re interested, write “article” in the comments and I’ll DM it to you.

@shai_s_shwartz Very cool benchmark, thanks for sharing. I am curious to know what's the comparable performance of humans, for example undergrads, grads, PhD level average performance, etc.

Are frontier AI models really capable of “PhD-level” reasoning? To answer this question, we introduce FormulaOne, a new reasoning benchmark of expert-level Dynamic Programming problems. We have curated a benchmark consisting of three tiers, in increasing complexity, which we call ‘shallow’, ‘deeper’, ‘deepest’. The results are remarkable: - On the ‘shallow’ tier, top models reach performance of 50%-70%, indicating that the models are familiar with the subject matter. - On ‘deeper’, Grok 4, Gemini-Pro, o3-Pro, Opus-4 all solve at most 1/100 problems. GPT-5 Pro is significantly better, but still solves only 4/100 problems. - On ‘deepest’, all models collapse to 0% success rate. 🧵

@YiTayML Not unexpected

Gemini 🙌

Do you remember when you joined X? I do! #MyXAnniversary

@_jpacifico @huggingface @ClementDelangue @julien_c @MistralAI @SebastienBubeck @ArashRahnamaPhD @jie_bing @natolambert @microsoftfrance Then, how long patience it took 😁?

@_hdarabi @huggingface @ClementDelangue @julien_c @MistralAI @SebastienBubeck @ArashRahnamaPhD @jie_bing @natolambert @microsoftfrance Thanks @_hdarabi Not too expensive , LoRA adapters on a single A100. The real cost, if you ask me: patience and determination ;)

My post-trained 14B model is now #1 on the French gov «Bac» benchmark, built from real national exam questions, ahead of DeepSeek-R1 70B, Mistral Large, Llama 3.3 & more. Started from the Phi-4 base model — model merging + DPO made the difference. Scale isn’t enough. Post-training is the key (right @maximelabonne ?😉)

@stoxplot 51 Trillion? 😱 What's the unit?

Samsung Q1 ’25: sales 79.1 T KRW (+10 % YoY). Phones/TVs (DX) still rule at 51.7 T, while chips/foundry (DS) add 25.1 T (+9 %). 36 % gross margin prints 28.1 T. Opex 21.4 T → op profit 6.7 T (8 %). Net hits 8.2 T (10 %). R&D 9 T—11 % of revenue. 👉Follow for more!