mceleri

I love local AI, but it just can't compete with frontier models (unless your "local" is something similar to a home datacenter).

"Let's assume the bulk of your one-hour session was with an average of 32 background agents doing different tasks, at an average of 20k tok/s input"

Then you do need 32 parallel sequences for your llamacpp (or similar), 20k context (which can be a bit small for a coding agent) * 32 parallel = 640,000. That is a big KV cache, even if quantized.

And what about the model itself? Local ones are impressive, but the frontier ones are a different game. (Again, unless you can run a very big model in your local setup).

Local AI is great for vertical tasks, NLP, and a lot of "simple" cases. Swarming it consumes the memory (parallel config). Swarm it with coding agents... I just don't see it working well.

Hi, looks nice. Do you have some benchmarks about which local model performs better with AlbertCode?

I like the idea to consider the token cost for the candidate, good. But why not keep it "on paper", asking how the candidate would use an LLM or agent and why? Actually using an agent means asking the candidate to deliver something that you must validate. And you then need to evaluate the candidate by looking at the output of the LLM indirectly. Isn't it better to directly ask the candidate to explain how an LLM could be used (or if/when it should be used) for your use cases?

At the end of the day, it is the same old 'interview coding game' problem, and I suspect it is working even worse now, using LLMs and agents.

Antigravity app, that seems just a wrapped browser. And yes, the browser as client of the deployed apps.

Hi, very interesting but are you sharing some technical details somewhere?

It looks to me that most engineers are between the AI entusiasts, not Luddites. We love to use it, we build tools around it, trying to optimize the whole chain... And so on, as always. I think your reasoning is more for startups and software houses than for engineers, because if the software cost go to "zero", a single engineer still have something to do. A company can't survive without good margins.

Well... Elliptic curve cryptography is also used in many other systems and connections, so yes, we would all notice if a mathematical trick broke it.

Dev guy here, so tailscale + Antigravity (+ GitHub). I use it to vibecode from my smartphone

I can't find any tech details, can we find some code/paper/docs somewhere?