Oz

@pinunhamon Not easy but possible

@oznova_ So they are easy to learn but hard to teach??

Yup. You can learn these things too, but frankly they may be as hard to teach to a human as to build into an RL environment

This feels so adversarial. Better option IMHO is to allow agents but do one of: - Have higher standards, see e.g. the CMU databases comp - Make grade contingent on a 5 min verbal defense - Allow custom agent that ships logs back for grading

@vboykis Honestly the thing I keep going back to is that the best engineers do some of their best work in a hammock (or a Google doc). If I focus on the big picture more than daily "productivity" I tend to let myself use agents to build insight rather than just features

@oznova_ These are great ideas! I’ve tried LLM as judge and writing myself to start and having the agent finish, but it feels really easy to lapse into passive generation. I’ll try the two options one next. Any other thoughts on introducing friction intentionally?

Something happens to my brain after agentic coding that I can’t describe. It’s like cognitive offloading which folks have already written about, but even more. It feels like I can’t think through problems anymore. Like a fog. Using agentic but losing my hard-won agency.

@jbmilgrom Yeah that makes sense too. I think it may come down to how much you can plan a priori. If you have a have a good map of the territory you probably want longer autonomous execution, whereas I often find I need to refine the map as I explore

interesting, how do you approach planning? my workflow is: 1. plan, plan, plan (can take a while or be quick depending on thing) 2. --dangerously.... (let it lose) 3. /pr 4. /code-style-review, /temporal-review,... (we have all these review skills for our stack) 5. look at code and ask questions to make sure good and i understand 6. back to 1 if necessary, sometimes even throwing out if i learned a better way, etc and could have many of these loops going on at the same time in different git worktrees for me, asking questions during 2 can be really harmful to productivity... im hoping agent makes best judgement so can get the full canvas of code on the board and is not sitting idle waiting for me to come back to that git tree for some reason

nevermind, 4.7 is trash indecisiveness at every turn is unbearable, questions of should i do A or B, when the answer is always, yes, keep going ugh town, USA. Back on 4.6, my workhorse

@vboykis Curious if you've tried anything else, like writing or rewriting portions yourself, or asking for two options for you to choose between?

The only things that I’ve found counteract this is putting the tool away and reading hard technical books and writing things on paper. I don’t think I ever felt like this when I searched Google or Stackoverflow for answers.

@odysseus0z Yes but not because of more visual explanations... It's because I can ask it to challenge me to implement the paper, giving me enough scaffolding/hints etc to keep me in the zone of proximal development

It is really the golden age for learners

@dreamwieber I've met *a lot* of people who first learned by typing games from magazines. It's like Hunter S. Thompson typing out Hemmingway

When I was a kid I went over a friend's house on a Saturday, and we hand-copied the source code for Lemonade Stand out of some magazine. And miraculously it worked. Not exactly 'coding', but just ... quaint, lol.

@epichrisis I don't remember much from that age but I can still picture my first QBASIC animation like it was yesterday

@oznova_ Ok but… you remember waking up early on a Saturday morning and having BASIC to yourself, complete with its godawful debugger, and whole universe to explore? Maybe that was just me, but I’ve been doing frontier work for 15yrs & still chase that feeling :)

Pretty crazy that I first learned to code by reading GORILLA.BAS (many years before Google, StackOverflow etc) whereas now when I want to learn something new I can ask an AI to build me custom progressive challenges

@tszzl Do you not think that AI research is mostly about having good, novel ideas? The legible technical skills amenable to RLVR seem to be a minor part of the job, no?

ironically think it’ll be a sad time for ai researchers this year. they are first in the hotpath of RSI and probably the market for them will shrink or at least their pricing power will be reduced as this generation of models commoditizes the skills that made them rare

@DavidPlakon This is going to be very confusing for me

Feature or bug in Warp? Let us know. Oz can help fix them on twitter now.

@turbochardo Students are active! I'm taking a bit of a pause from new videos while I figure out if I want the next series to involve more agent-based coding. There's a lot of good existing content though!

@oznova_ are you still releasing new content on csprimer / is the community active?

Yes! In addition: - Schools have barely changed in response to computers and the internet, so will do a poor job with AI - Focus is entirely on highly legible versions of "subjects" (math, english etc) when there is more value in what is in the gaps - There is basically no attempt to develop what is important (self understanding, virtue, appreciation of how the world is built etc) - The grade based school system keeps siblings unnaturally separated from one another (let alone their parents) - In many school districts, it's hard to pull kids out for long stretches for travel etc

Toxic and out-of-date education practices that will likely make me a home-school parent (private schools are not immune from these issues): ➡️age- and geo- limiting learning cohorts ➡️"scheduling" learning by periods in a day and limiting content by grade "level" ➡️treating everyone in a class with the same processes/standards (like a doctor giving every patient the same medicine) ➡️lack of accountability for the student (grade inflation, pass/fail), and teacher (tenure, unionized pay) ➡️lack of feedback loops with the "real world" ➡️lack of personalization ➡️high failure rates (even the "best" high schools have a 90% failure rate for colleges I hope my child attends -- while my business helps almost all students achieve a top 30 college, 50+% hit Top 10 and Ivy League, and 30% hit Harvard, Yale, Princeton, Stanford or MIT) Am I missing any other big reasons?

@nwilliams030 Kitchen Table Math is a nice map of the territory of early math education. Nebel's Elementary Education might be an interesting counterbalance to Well-Trained Mind in that they're both broad overviews but Bauer argues you can't really do science yet whereas Nebel disagrees

absolute favorite reading related to homeschooling in the young years? 3-5yrs? so far we have - The well trained mind by Susan wise Bauer (ofc) - laying down the rails (great on virtue instillation in the home) - a year of forest school - Singapore math 1A

@DJCrabhat @fkasummer Maybe there's a third category of other/esoteric :)

@oznova_ @fkasummer I think we miss the richness when we talk only of imperative languages here? There's a whole world of declarative languages that I think are much harder to jump around between. Everyone's a polyglot until they get slapped with a monad 😁

The answer to "Do you know programming language X?" is always yes if you can already program in ANY language. If not, you didn't really learn to program.

@amasad Congrats man! Strength to strength

We’ve raised $400M at a $9B valuation. Investors include Georgian, G Squared, Prysm, 1789, YC, Coatue, a16z, Craft, and QIA, with strategic investments from Accenture, Databricks, Okta, and Tether. We’re also lucky to have incredible individuals backing us, including Shaq and Jared Leto. This funding will help us scale our ambition and expand beyond coding into AI systems that center human creativity. Replit is now used at 85% of the Fortune 500. We have an opportunity to help shape the future of work. One where AI abstracts away the boring parts and humans shine as creative directors. We’re also investing more globally, particularly in Europe, Asia, and the Middle East. Innovation can come from anywhere in the world, and we want to help unlock it.

@mboverell @jmj I spoke recently with a friend in PE who owns a large chain of dental practices... AI is very interesting to him!

@jmj Until we have robot arms you would let put things in your mouth or use scissors around your neck?

There's a very clear divide right now in my friend group between people who are AI optimists & pessimists. The pessimists are very scared right now. They wake up in fear that their jobs and livelihood will go away in the next few years. Most of them work in tech, but are a bit later in their careers. I can see the paralysis on their face. They say they're using AI tools, but they're not really making all that big of an effort. It's almost like seeing people who couldn't function during COVID becoming victims to the changing world. /// The optimists think that this is the best time ever to be working in technology. If today is Super Bowl for builders, they are the ones who are playing the game. They don't wake up in fear because they're having too much fun. I see them building software, running experiments, and trying to figure out what's next. /// Are you an AI optimist or pessimist? I can tell you that's much more fun to be an optimist right now. Don't let yourself get in your own head. AI will destroy your confidence if you let that happen. Embrace the unique moment we're in, mostly because there's no going back. And it's the best way to live.