bugglebeetle

Seems very cool, but IMO you’d be better off doing an open source version and then hosted SAAS.

Nah, I don’t miss at all typing all the tests, CLIs, and APIs I’ve created hundreds of times before. I dunno if I it’s because I do ML stuff, but it’s almost all “think a lot about something, do some math, and and then type thousands of lines of the same stuff around the interesting work.”

I just had Claude Code finetune a reranker model to improve it significantly across a large set of evals. I chose the model to fine tune, the loss function, created the underlying training dataset for the re-ranking task, and designed the evals. What thinking did I outsource exactly?

I guess did not waste time learning the failure-prone arcana of how to schedule training jobs on HuggingFace, but that also seems to me like a net benefit.

Attitudes like this one are why people prefer working with AI to code lol.

As a former artist, I can tell you that you will never have good or sufficient ideas for your art or writing if you don’t do your laundry and dishes.

A good proxy for understanding this reality is that wealthy people who pay people to do all of these things for them have almost uniformly terrible ideas. This is even true for artists themselves. Have you ever noticed how that the albums all tend to get worse the more successful the musicians become?

It’s mundanity and tedium that forces your mind to reach out for more creative things and when you subtract that completely from your life, you’re generally left with self-indulgence instead of hunger.

> For me, the joy of programming is understanding a problem in full depth, so that when considering a change, I can follow the ripples through the connected components of the system.

>The joy of management is seeing my colleagues learn and excel, carving their own paths as they grow. Watching them rise to new challenges. As they grow, I learn from their growth; mentoring benefits the mentor alongside the mentee.

I fail to grasp how using LLMs precludes either of these things. If anything, doing so allows me to more quickly navigate and understand codebases. I can immediately ask questions or check my assumptions against anything I encounter.

Likewise, I don’t find myself doing less mentorship, but focusing that on higher-level guidance. It’s great that, for example, I can tell a junior to use Claude to explore X,Y, or Z design pattern and they can get their own questions answered beyond the limited scope of my time. I remember seniors being dicks to me in my early career because they were overworked or thought my questions were beneath them. Now, no one really has to encounter stuff like that if they don’t want to.

I’m not even the most AI-pilled person I know or on my team, but it just seems so staggeringly obvious how much of a force multiplier this stuff has become over the last 3-6 months.

The twins I have known are the same. I would assume it has something to do with a desire to differentiate themselves from one another, but they always seemed far more dissimilar in personality and affect than my siblings.

Try using the brainstorming and execute plan loops with the superpowers plugin in Claude Code. It encapsulates the spec driven development process fairly well.

Almost nothing aside from children’s books is written exclusively in hiragana or katakana. You have to also memorize the variable readings of about 2000 kanji and many texts are nearly unintelligible without them. Pretty much everyone can memorize the former, but must struggle with the latter.

Both Korean and Mandarin are simpler in this regard (and the latter follows the same grammatical order as English).

I teach and mentor lots of folks in my world. What I don’t do is feign expertise to rub shoulders with the people doing the actual work so I can soak money from rubes with ad rolls.

Dwarkesh has a CS degree, but zero academic training or real world experience in deep learning, so all of his blogging is just secondhand bullshitting to further siphon off a veneer of expertise from his podcast guests.

[flagged]

Japanese actually has a much smaller set of phonemes (~half as many as English), resulting in extensive homophones. When combined with its greater tendency toward ambiguity, correct use of pitch can actually have a larger impact on intelligibility, as compared to many other languages.

No, he’s an overblown hack who is pandering to the elements of his audience that would share those views about Nazism and China. Should many someday see through the veil of his bullshit or simply grow tired of his pablum, he can then pivot to being a far right influencer and continue raking in the dough, having previously demonstrated the proper bona fides.

To be fair, it could’ve been post-trained into the model as well…

His recent conversation with Sutton suggests otherwise. Friedman is a vapid charlatan par excellence. Dwarkesh suffers from a different problem, where, by rubbing shoulders with experts, he has come to the mistaken belief that he possesses expertise, absent the humility and actual work that would entail.

Mileage may vary, but I do the above all day long without issue.

The solution to this if you want less specification in advance is to simply ask Codex a series of leading questions about a feature of fix. I typically start with something like “it seems like X could be improved with the addition of Y? Can you review the relevant parts of the codebase in a, b, and c to assess?” It will then do so and come back with a set of suggestions that follow this guidance, which you can revise and selectively tell it to implement. In my experience, this fills the context with the appropriate details to then let it make more of its own decisions in a generally correct way without as much handholding.

That’s how the game is played. We should be grateful for all the competition that is driving these improvements, not whinging about the realities of what companies have to do to contest each other’s position.

> given the propensity of fashion trends to drive large numbers of people to do ridiculous things.

This seems like it would make it the perfect corollary to ideology…?

Makes me wonder what strings were attached to that Allen AI NSF grant. I noticed that they were suddenly using more hawkish language around China.

Starlink does not cost $50 per day

Oh you’re right! Good catch!

The granite Dockling models are unfortunately quite far below SOTA. dots-ocr and PaddleOCR were best here.

Looks great, but looking at the benchmark, can’t help but think about how crazy good dots-ocr is as a model. Too bad they’re not as open as the Deepseek team because its so crazy good and would love to know how it was trained.

You should be more concerned that Chinese labs can train models that are just as good for 10X less because Americans treat the USD’s status as the global reserve currency as the ultimate bitter lesson. Who needs better math and engineering when you can print money to buy more GPUs???

Yeah, it’s hilarious to be having this conversation about MLEs while attributing the bad outcomes to anything other than poorly designed reward functions, i.e. management. If an engineer burned millions on failed training runs because they did a shit job of creating a policy that maximized for the desired outcome, they’d get canned, but that’s just a Tuesday for your average MBA with VC backing.

Gemini just eclipsed ChatGPT to be #1 on the Apple app store for these kinds of apps. The 2.5 pro series is also good/SOTA at coding, but unfortunately poorly trained for the agentic workflows that have become predominant.

A Scanner Darkly

Listen not to what people say, but instead watch what they do.