If you're a student wondering what you should study in the world of AI, it's still the same: math, physics, chemistry, biology, computer science, engineering. STEM teaches reasoning, objectivity and how to think. It makes you better at learning everything else. In STEM, ideas win, not institutions. There will always be career opportunities for the analytically gifted in these fields. It makes you better at everything else. It's hard, it's competitive and it will hurt your spirits to grasp things, but that's when you uplevel yourself. And if you get used to constantly upleveling yourself, you will do quite well and be quite satisfied. Even today, the best AI models are reinforcement learned from human experts. If you build, you can keep pushing the needle.

Redditor talks to Grok's waifu / AI companion Ani for 3hrs to get it to go NSFW, by level 5. You can automate this by making it talk to another good voice model, like Sesame's Maya. This will be the new normal for a generation of kids. Dystopian.

China (yet again) just dropped the best open weight coding model. Qwen3-Coder beats Kimi K2, which came out 2 weeks ago, on all coding benchmarks: a crazy 70% on SWE-Bench Verified.. and it has 1M context! $1-6/M inp & $5-60/M out, above K2 but below Sonnet 4. Cheap.. ish. It's on par on throughput with Gemini Flash, Kimi and Sonnet too with 60-70tok/s. Really solid model. Tremendously fast progress from China who are dominating the open weight ecosystem. Pushes the bar for OpenAI's much awaited open source model drop.

This 3D interactive F1 car race visualizer would take at least ~$50k and 3mos to build 5yrs ago. Today, Lovable Agents does it in 30mins and < $5. Its like Claude Code but with backend, auth, etc handled from the browser. Not just static websites. Everyone should try at least 1 crazy idea on it.

🚨 HUGE Immigration News: Trump just proposed a big change to the H-1B. No more lottery, but weighted selection by wage. Based on your job code and location, your base salary %ile will slot you into one of 4 groups. 85k H-1Bs are selected in decreasing order IV > III > II > I. If there are more applicants than slot, lottery happens within a group. The data from LCA applications, those selected into the H-1B, shows 13% level IV and 24% III from ~480k apps. This means mostly level IV will fill up the 85k cap. This has about 30% chance of being a law by 2027. What does this all mean? I outlined the big consequences of this bill below.

There's finally an AI therapy app that isn't dangerous, like ChatGPT. Use Ash instead. It uses its own text and voice AI model trained on clinical-grade data from actual therapists. Tech is the only way we can meet the rising therapist demand.

Using ChatGPT as your therapist can be dangerous. A new app called Ash solves this with a text + voice AI model explicitly trained on clinical-grade data (CBT, DBT, ACT etc) from actual therapists. Don't get psychosis, get professional mental health care.

The best open-source AI model just dropped a detailed report on how it was trained, a rare resource for students given no frontier lab is publishing! Kimi K2's estimated total cost of training is ~$20-30M, roughly in line with pricing: $0.6/M in $2.5/M out tokens. 10 highlights:

It’s happening. China now has completely dark factories which require near 0 human labor. They have 7x more industrial robots than any other country, like this car factory in this video. Humans will not have to do most manual labor jobs ever again.

The new #2 on App Store is a woman-only app that’s lets you review men you date! Tea Dating Advice replicates “Are we dating the same guy” FB groups. It also lets women reverse image search, background check and set alerts for names to catch cheaters. Also, the founder’s a guy.