I think this is the wrong path in LLM and SWE optimizations:
1) Programming language training happens by volume, and the amount of JS/TS/python out there, and the rate it's growing at - is causing a training effects loop, which means for a few generations of models, these will be the best performing languages. Will be hard for a contender to spin up.
2) At some point, if we plateau on productivity - then efficiency improvements will happen, which will open a door for programming languages that maintains productivity, but is 10x cheaper on cost.
3) I think more immediate gains are at the cloud level. IMO, one of the reasons Google cloud is performing better(along with firebase) is much better overall CLI experience, leading to a pleasurable experience developing against it. This part of the market is ripe - whoever builds a most LLM friendly cloud has a shot of shooting up. Hence projects like exe.dev, and whatever cloudflare and vercel are trying. It would be good to have some shakeup in the cloud world.
I thought it would be an informative discussion to have to apply these rubrics to LLM and AI use coming in across domains, in the context of Freedom of Refusal / Reversal.
Crystal was always something I looked at ~5 yrs ago - alongside golang and rust. But it would out of the new cycles. Looks like they solved for 2 big things in these years: fast development cycles and windows support.
Very cool. I wonder how it stacks up against golang for production apps. Anyone comment?
If your concern is someone selling GoModel as a service, you could add a license provision for that. Technically it'd no longer be open source, I think, but most people won't care.
Hi Eric, would like to understand how you approach that steering. It's a problem statement I've been working on as well, would like to compare notes. Couldn't find your contact - mine is in my profile.
I think "articulator" is a better term than "incanter" - specially with the recent SOTA models. The previous ones suffered from things like - special magic phrases improving performance on benchmarks, but this has gotten significantly reduced - can't remember the last time I used those hacks. Right now the skill is clearly explaining in plain English your needs.
Which means clarity of purpose, and then articulating it to LLM is the skill. English + critical thinking.
it's a novelty to see the connections. One way it will be useful is to connect every character to the stories they're part of - either in the site, or in new tab. this will allow exploring the stories for each one of them. This will make people come to this for more than novelty, imo.
I understand people have a viewpoint here about not giving time to large behemoths. I'll counter with a story and perhaps a larger point.
Back in 2006/7 I had an idea for a project for which, in all enthusiasm, I setup a mailing list, but ended up never pursuing it. It's a very unique name.
In 2012, another developer landed on the same name for their project, but saw that the mailing list was taken up and reach out inquiring if he could take over, and I obliged because here's another person doing something in cryptography and open source, 2 of my favorite things then (and now).
The project was "scrypt" and the developer was Colin! :) I knew nothing about Colin or tarsnap then, IIRC.
Sometimes you just do kindnesses of which you're able, with people who you feel a sense of community with, without expectation of anything commercial. Karma adds up, and it's benefits are large, though hard to always articulate.
they but "minimal dump DRM" into their client (supposedly, from people which leaked the linked source code, no me)
easy to circumvent
but would fall under "circumventing security protections"/"hacking their API"/etc. And due to the sometimes very unreasonable laws the US has in that area they can use that to go after anyone providing a workaround.
Through that maybe won't work well for the EU, I'm not sure how much the laws have been undermined in recent years but we had laws which made it explicitly legal to circumvent DRM iff it's for the sake of producing compatibility (with some caveats).
I think the law just says that it's legal to circumvent DRM for compatibility - they don't define DRM or compatibility. It's one of those vague laws that you only know if it matters when it gets tested in court.
I think if you were going to send the same harness/prompt traffic as Claude Code, then you’d just use Claude Code. Alternatives generally are trying to do something different, thus are going to be easy to detect.
The best way to learn to build an agent is to learn and use pi.dev . The homepage is a masterclass of explaining the main loop with 4 tools (read write edit bash) and how they enable any flow. Skills and Agents.md on how the agent can be guided.
If you understand the above, you're 80% of the way to building any purpose agent. System prompt, project prompt, and the tool call loop.
1) Programming language training happens by volume, and the amount of JS/TS/python out there, and the rate it's growing at - is causing a training effects loop, which means for a few generations of models, these will be the best performing languages. Will be hard for a contender to spin up.
2) At some point, if we plateau on productivity - then efficiency improvements will happen, which will open a door for programming languages that maintains productivity, but is 10x cheaper on cost.
3) I think more immediate gains are at the cloud level. IMO, one of the reasons Google cloud is performing better(along with firebase) is much better overall CLI experience, leading to a pleasurable experience developing against it. This part of the market is ripe - whoever builds a most LLM friendly cloud has a shot of shooting up. Hence projects like exe.dev, and whatever cloudflare and vercel are trying. It would be good to have some shakeup in the cloud world.
Anyway, this is where my thoughts are currently.
reply