This is why open source matters so much! When community-driven tools like @cline can give every developer free access to ~300 tok/s and 1M context, the barrier to building truly drops to zero. No gatekeeping, no massive API bills - just pure enablement for builders everywhere (especially students). Go create something cool! 🛠️✨

Nope. It's still true. Where is your domestic robot? Where is your Level-5 self-driving car? Where is your robot car that can learn to drive in a few hours of practice like any 17 year old? Where is your AI system that can understand the real world and quickly learn new skills like a house cat? There is no question that AI will eventually become as intelligent as humans in all domains. *** BUT *** 1. We're still far from that, even if AI and computer technology surpasses humans in an ever-increasing number of tasks. 2. It won't be based on LLMs, although LLMs will have a role to play (e.g. as a text interface).

If you're entering college or know someone who is, PLEASE know that a degree doesn't magically grant you a job (and it hasn't for a long time) At a minimum, you should have some form of professional experience every summer, working your way up, like: Freshman year - 10 hours a week at an unpaid internship that will teach you office skills, 30 hours a week at whatever job pays (landscaping, lifeguarding, etc.). On your resume, only list the unpaid internship for that summer, let people assume it was more than 10 hours (but obvs be honest if they ask). During Freshman year, get some campus leadership roles, TA-type positions, and things like that as applicable. Put yourself out there, people should know you! Sophomore year - Try to get into a more serious internship, likely also unpaid (sorry) unless you're in STEM or business/finance. This one should be more of a name brand that people in your region would know. Supplement with other work as needed like you did last year. During your sophomore and junior years, also take on projects for professors and departments. This spreads out the people who can be references and gives you broader experience to reference in interviews. Junior year - start networking for this EARLY, like a year+ early, it is the main one that matters. Ideally this will be paid, but the most important part is that it's somewhere that could hire you/a strong enough brand that it could get you hired elsewhere. Senior year - the ideal situation is a return offer from your junior year placement. If not, rotational programs, analyst programs, and the like are your friend (and typically involve very early applications) None of this is set in stone, but the main thing you need to avoid is an empty resume by senior year That is how a lot of people are entering the job market and it is absolutely brutal

Billing paper cut fix in @GoogleAIStudio 💸 You can now see all your eligible GCP promotional credits for Gemini API without leaving the AI Studio Billing page. Don't forget to redeem your $10-100/month credit in AI Studio when you have a Pro or Ultra subscription and a Google Developer Program profile!

I will admit, I am enjoying vibe coding. I may have been a teensy tiny bit wrong.

MiMo-V2.6 is "simply" the best (for now). Despite its simple architecture design it's currently No.1 in the open-weight benchmarks (weighted average). With "simple," I mean a classic Grouped Query Attention (GQA) with Sliding Window Attention (SWA) at a tiny 128-token window size. So, that underlines one of the points I've been trying to make in recent months: most of the progress still comes from the data and post-training recipe improvements. Fancy attention variants are just mostly efficiency tweaks. What are some of the training data improvements and recipe improvements? The MiMo team shared a pretty detailed technical report. Lots to carefully digest there, but in short, there are a few things that stood out: 1. An increase in agent tasks; also training across different harnesses (the average DeepSWE pass@1 accuracy on held-out harnesses improved from approximately 50% -> 66%). 2. Better reward signals: they replaced a simple correctness verifier with an agentic grader that looks at the execution traces as well. 3. Large RL batches (1,568 prompts × 16 rollouts = 25,088 trajectories) and 2.7–3.7 billion training tokens per update (unclear, though, what the predecessor used).