It isn’t about intelligence per se. It’s about computation. We have now the ability to just ask for a computation to be done for us. Literally, just ask in plain English. Computation itself is the intelligent thing here. Before we needed to program it. Now weights inside an LLM

The post-patience era for software engineers is upon us. Patient experts were our last refuge to learn mastery. Masters now may just send us away with a prompt to try out with Codex.

while programming with AI, hard to decide whether to be inspired by how much it does, or annoyed by any manual work left for you to do

It’s not that you need a reason to be happy. It’s that you don’t desire it enough. If you desire it to the exclusion of everything else, then you will be happy. It’s a choice, as @naval says.

FSD is a modern miracle. If you haven’t experienced it, a test drive won’t give you the full experience. Rent a Tesla for a month. Start using FSD and you’ll know that you live in a parallel world. It’s very very hard to describe because it’s one of those things like

Great work by @AbsBhandarkar to build a system to detect jailbreak attempts and prevent it from being submitted to LLMs.

Eric Schmidt should be required listening.

Identifying queries that may induce a jailbreak is an interesting problem to solve in AI security. Many solutions train language models on known query patterns. I believe we may need to solve it with a cascade of classifiers - rules, encoder only (BERT like) and LMs. Rules

Why are we waiting for AGI? Its already here and its called Claude Code.

Keep increasing the context window. We need this to tell Claude all the system tools and utilities so it can become Claude OS

Petty Corruption vs IQ Let’s say you’re at the top of the bureaucracy and solve for the toughest exam and yet - you still break your back every day on broken roads even if you’re chauffeur driven - you have to send your kids halfway across the world for a decent education and

Likes of Google never sleep. Decades later, systems like Search, Maps, Email, Docs, Browser, YouTube, Android and so many more keep on going. They weren’t sleeping, they just got worried about prestige. That happens when you run world-scale systems.

Great way to find the human users who remove emojis from the bio vs those who don’t. Those who don’t are people who didn’t see this post, don’t care or want to stick it to the algorithm. Bots are controlled by bot managers who are usually people. If autonomous agents are

The article talks about many companies trying to build in house and not seeing expected results. But when they buy, there is more success.

The real issue with knowledge cutoff in models powering advanced voice mode is you cannot riff with them on how to use o3 or GPT-5. They see still stuck for the most part in the 4o world. The best AI you can talk to should also be the AI that has the most current knowledge of

Glad to hear this from Jeremy Howard. I’ve always felt the same way about jQuery and have continued to use it.

For many years, I've been looking for the perfect replacement for @jQuery. Modern vanilla js has many of the features that made jQuery popular, but often with worse ergonomics. Today, I finally discovered the solution: Just use jQuery. :D

👨‍🔬 LLMs are your copilot. RDKit is your reactor. SMILES is your language. Vector DB is your memory. https://t.co/IYS1eLj3Cw is your blueprint. Now go build the terminal for your niche.

The best part? You don’t need a big team. Just: - Good data - A small LLM budget - Domain intuition - A clean UI Feeds for chemists. Alerts for safety. Search for scientists. Built by you. For your field. Start with search. Grow into information feeds.

LLMs didn’t just make search smarter, they made vertical, expert-driven search viable. Feeds, search, alerts - all fine-tuned for chemistry, not clickbait. The future belongs to niche platforms built by insiders, not influencers.