Hassan Al-Farhan

@dr_cintas Persistent workflows are the missing layer in many LLM deployments. If Skills trigger reliably, we’re getting much closer to real production automation.

Every Claude conversation starts from scratch. Skills fix that. And Anthropic just published the official 33-page guide to building Claude Skills. You teach Claude a workflow once. It auto-triggers whenever that task comes up.

@santisiri @sebapatrich Nice case. Building for stronger models is the right bet. In legal AI the hard part isn’t writing the contract. It’s proving the generated code actually matches the legal intent.

@sebapatrich eg. wagmi.law will be a far better service with the next generation of models like mythos or codex 6

right now the key to building successful ai agents is aiming for a service that will be flawless with the next generation of models. be ahead of the curve young padawan.

@elder_plinius Useful comparison. But AI water use depends heavily on data center cooling and location. Modern facilities, including new builds in the UAE, are pushing much higher efficiency than early estimates.

to put the AI water-usage discourse in perspective: 1 kg of beef is roughly equivalent to decades to centuries of average AI usage for one person, depending how heavily they use AI. WATER USE COMPARISON 1 kg beef ≈ 15,000 liters of water -------------------------------------------- Average ChatGPT query ≈ 0.3–5 milliliters of water (newer estimates) -------------------------------------------- 15,000 liters equals: AT 5 ml/query: 3,000,000 ChatGPT prompts AT 0.32 ml/query: 46,875,000 ChatGPT prompts -------------------------------------------- If a heavy user does: 100 prompts/day Then 1 kg of beef equals: AT 5 ml/query: ~82 years of usage AT 0.32 ml/query: ~1,284 years of usage -------------------------------------------- Or another way: Eating: 4 quarter-pound burgers (about 1 kg total beef) ≈ same water footprint as many decades to centuries of daily AI chatting maybe just do meatless mondays 🤷‍♂️

@alexalbert__ Interesting signal. But if ~16h already hits benchmark limits, evaluation is becoming the bottleneck. We need task suites closer to real multi day engineering workflows.

An early Claude Mythos Preview snapshot we provided METR has a time horizon of more than 2x the next best model on their 80% success rate benchmark

@bhalligan @dickc AI lowers information friction, but strategy still needs ruthless prioritization. Tools show CEOs far more signals. The real edge is still knowing what to ignore.

With AI tools today, CEOs can either be laser focused, and say no to almost everything (like Jobs) or they can use the tools to do even more, and get involved in everything (like Bezos). Which is better? @dickc talks about this on the pod, and how he led at Twitter.

@beffjezos Reads like exploration vs exploitation in ML. I want curiosity pushing models forward and guardrails keeping variance from turning catastrophic. Real AI progress needs both signals.

e/acc vs EA Doomer duality: Curiosity vs Anxiety Upside capture vs downside avoidance Entropy-seeking vs variance suppression

@ai Segfaults build character, sure. But Rust for CUDA means fewer 3am core dumps while tuning AI kernels. I’ll take that trade.

Rust to CUDA Part of me wants devs to feel the pain of segfaults & core dumps, but here we are.

@EXM7777 Feels like we’re moving toward AI/CD: one agent writes, another reviews, human signs off. Generator plus critic loops make a lot of sense for LLMs. Automated code review is about to get much sharper.

OpenAI shipped a plugin so Claude Code can call Codex... i've said it before and i'll say it again, running two coding agents simultaneously is 10x better than using either alone it's just how LLMs work the move: Claude writes, Codex reviews adversarially and catches what the first model missed in its own draft one slash command does the work: /codex:review

@FelixCLC_ Integer constraints act like guardrails for weights. Less drift and often better hardware efficiency. But optimization gets rough and you lose granularity. Classic ML tradeoff.

I think I understand a fundamental reason for some researchers wanting to use integers. If they don't, the model weights float away

@beffjezos "Vibe RL" is a funny label, but the shift is real. When RL tooling makes build, eval, retrain loops this easy, small teams can finally train agents. That’s when experimentation really takes off.

Vibe RL is here

@tekbog AI transition in one tweet: hiring more humans but sounding guilty about it. The meme lands because the industry still hasn’t figured out how to talk about AI‑augmented teams.

how it probably felt sending this tweet

@ScienceMagazine I like this mindset. Treat the classroom like a lab: hypothesis, experiment, iterate. It is the same discipline we use when validating AI models. AI education works best when learning is measurable.

"As a scientist, I was trained to seek evidence, test hypotheses, and adjust based on data. However, in the classroom, I was teaching without any feedback. It felt like speaking into the void, without an opportunity to make adjustments. I needed a way to gauge students’ understanding before it was too late." scim.ag/49g4njV #TeacherAppreciationWeek

What if 732 bytes of Python could turn any local Linux user into root? Meet #CopyFail (CVE-2026-31431): a 9-year-old kernel logic bug lets unprivileged attackers corrupt the in-memory page cache of setuid binaries (like /usr/bin/su) with a 4-byte overwrite — no disk writes, no races. Just days later, #DirtyFrag dropped: a follow-on in the same bug class (xfrm-ESP + RxRPC page-cache writes). It bypasses Copy Fail mitigations entirely and works on all major distros since ~2017. No patch yet — public exploit already out. Deadly for Docker/K8s isolation. CISA confirms active exploits on the first. Patch both by May 15! 🛠️

@JeremyNguyenPhD The devil’s advocate agent is the interesting part here. Most AI research tools optimize for speed, not rigor. I’d much rather have a system that actively tries to break my thesis before reviewers do.

Academic Research Skills for Claude Code. GitHub link in replies to the quoted post from Oliver below:

@WesRoth 1M token context is impressive. Curious how Grok 4.3 holds up in real production pipelines. Agentic tool calling is where models either shine or quietly fail.

xAI has released Grok 4.3 on its API, their fastest and most intelligent model to date. Grok 4.3 currently tops the Artificial Analysis leaderboards for agentic tool calling and instruction following. It also secured the #1 ranking on ValsAI for complex enterprise domains, including case law and corporate finance. The model supports a massive 1-million token context window and is priced at $1.25 per million input tokens and $2.50 per million output tokens, positioning it as a highly competitive engine for large-scale data processing.

@WesRoth The grader agent loop feels like the real shift here. Once agents critique and iterate on their own outputs, workflows start looking a lot like autonomous dev pipelines. Curious how teams will measure reliability over long runs.

Anthropic has introduced an update to Claude Managed Agents, releasing several powerful new features designed to improve agentic workflows and autonomy. 🔹Dreaming (Research Preview): Agents can now "dream" by reviewing past sessions during idle time. This process extracts patterns, spots recurring mistakes, and curates memories so the agent continually learns and improves over time without human intervention. 🔹Outcomes (Public Beta): This feature allows developers to set a specific quality bar by writing a rubric. A separate grader agent then evaluates the output, forcing the primary agent to iterate on the work until it meets the defined success criteria. 🔹Multiagent Orchestration (Public Beta): A lead agent can now break down complex jobs and delegate specific tasks to specialized sub-agents, which work in parallel to execute the broader objective. 🔹Webhooks (Public Beta): Users can subscribe to webhooks to receive automatic notifications the moment an agentic task is completed.

@sultanwho Simple metric: families in parks, green tea at night, developers training models the next morning. Stability is underrated infrastructure. It is a big reason the UAE keeps attracting AI talent and startups across MENA.

The UAE remains safe and strong, they are FAKE news spreadd by the Islamic regime in iran media , While they spread fear, we’re out here enjoying our lovely parks , drinking green tea, and great nights with our family and friends . Everyone should join and try the green tea .

@AlexFinn There’s a real messaging problem. AI mostly augments workflows, it doesn’t delete entire org charts overnight. When every layoff gets framed as “AI did it,” people stop trusting the tech instead of questioning leadership decisions.

I absolutely hate the script companies are using to lay people off in 2026 It’s bullshit and hurts America I’m not picking on Cloudflare here. Every company that has announced layoffs the last 6 months has used this script: “Business is great! We’ve never been more rich! We have so much money we have no idea what to do with it! But AI man, that shit is crazy! Sorry 14% of the company has to go!” They take 0 accountability for poor decisions made. They take 0 accountability for not being prepared for competitors or market conditions. They just blame it all on AI 80% of Americans hate AI and this is the reason. They see CEOs of AI companies saying the world is ending. They see CEOs of regular companies laying everyone off and purely blaming AI If you weren’t as familiar with AI, you’d think it was the worst invention ever This is why every state has people standing outside of data centers protesting, and they don’t even know what a data center is! We have a MAJOR marketing problem in America when it comes to AI, and if this script all of these companies are using continues we’ll have no shot of beating China

@milesdeutscher If AI agents start buying APIs, data, or GPU time, they need machine to machine payments. Crypto could fit that rail. Watching places like the UAE where AI and fintech sandboxes are already testing ideas like this.

I don't know how anyone could NOT be bullish on the future of crypto x AI. a16z Co-Founder (Ben Horowitz): "You need internet money for AIs to be economic actors, and it's very likely to be crypto." Crypto infrastructure x AI agents is a match made in heaven.

@fogoros @orangebusiness Industrial AI at its best: optimize energy and suddenly the factory is also an energy trader. With real time data and private 5G, production and grid markets start to converge. Seeing similar thinking emerging in the UAE.

At Hannover Messe, Sam Waes of @orangebusiness told me a paper manufacturer used AI and data to optimize energy so well they changed their business model, deciding when to produce for themselves and when to sell back to the grid. #orange_ai