Kamil Nicieja

PSA: SF Ruby is hosting an AI hackathon… and I’ll be there too! Rails has always been the go-to framework for solo developers. So here’s the big question: *exactly* how much further can we take Ruby specifically with LLMs? Let’s find out!

claude code is very very bad at writing tests and will always try to make the job easier for itself by cheating, BUT it’s very very good when it can iterate against a good set of tests, so perhaps this means tdd is the best way to program with an assistant

Wow, this is a pretty impressive idea.

Just came across DeepSeek—it really showcases the thought process of a reasoning AI, something OpenAI's o1 doesn't quite do. This got me thinking about test-time compute this week.

(This doesn't factor in potential decreases in inference costs over time, which could change the equation. No idea what’s the trend here.)

Just came across DeepSeek—it really showcases the thought process of a reasoning AI, something OpenAI's o1 doesn't quite do. This got me thinking about test-time compute this week.

Of course, not every query will take 60 seconds, even with test-time compute. So annual expenses might realistically fall between $9 billion and $48 billion. That's potentially 1.28 to 6.85 times the current demand. So where does that leave us?

Just came across DeepSeek—it really showcases the thought process of a reasoning AI, something OpenAI's o1 doesn't quite do. This got me thinking about test-time compute this week.

If 4o suddenly disappeared and o1 handled all interactions—with responses taking 12 to 60 seconds (as seen with DeepSeek, where longer seems better)—what would that mean for costs? Assuming inference costs increase linearly with response time, infrastructure expenses could rise by 3 to 15 times.

Just came across DeepSeek—it really showcases the thought process of a reasoning AI, something OpenAI's o1 doesn't quite do. This got me thinking about test-time compute this week.

I read somewhere (can't recall where) that inference alone might cost around $3 billion. Let's assume that's accurate for this discussion. Now, imagine if training new models stops entirely, leaving only inference costs. How would that impact expenses?

Just came across DeepSeek—it really showcases the thought process of a reasoning AI, something OpenAI's o1 doesn't quite do. This got me thinking about test-time compute this week.

Disclaimer: I don't have direct experience with AI lab budgets, so this is purely speculative. Everything in this thread could be inaccurate. If anyone can provide corrections or more accurate information, that would be greatly appreciated.

Re: Google having to sell Chrome. Many have criticized this decision, but for the first time in a while, there could be genuine buyers in the market who aren’t part of Big Tech like OpenAI or Perplexity. I’m not so sure this is a bad thing.

Testing writing tools from macOS 15.1. Hello, could you please confirm that Siri is now functioning correctly?

I swear that 80% of Interstellar’s cultural impact comes from its soundtrack and the way it visualizes black holes.

Wow, I just realized the December SF Ruby Meetup is only 18 days away. I'll be presenting Lammy, my lightweight LLM library for Ruby. Time to start working on the presentation and polishing the code to make sure it’s ready to show!