kolinko
I just don't care how it looks to a human - LLMs have their own criteria of what is readable and manageable to them, and if you try to enforce human readability you will bump into a ton of antipatterns that will cause more and more issues.
Also, the default code management tools from Codex/CC are either basic (Grep/Edit), or - if you try to use some open source ones - unlikely to be used by agents. It took some time, but once I got agents to navigate the codebase efficiently, and produce code that is accessible to them, I stopped having problems with bugs or increasing complexity.
If you begin treating human interventions as an error in a process - and do after action reviews and so on after each such intervention - then after a few months you end up with quite a smooth-working system.
And I've spent around $100-$120k in api calls this year (total, both providers). Terra is awesome (price-performance) on noncreative things, but I don't allow it to write anything besides one-off scripts. Astra is decent for adversarial review, but only when you watch closely for her to not derail project into directions that are away from goal but safe to build, and Sol I honestly don't know where I'd use.
Having said that I agree crawling should be legal in general. If an agent is acting in my name, it is not much different than a browser.
The biggest challenge is in an office, or other places where I want to maintain privacy. Ideas?
By "telling" vs "talking" I meant - I have a setup that is 95% noninteractive - oneshotting most tasks, with a decent verification setup and a good memory/hooks to be confident the system does what I want and is reliable.
When something off happens, I spin up an after action review session that pinpoints the guilty sessions, interviews them, the whole investigation, and either handles improvements in the process by itself (most of the cases), or asks me for feedback. For the last three months I don't think I had a place where my input was really needed. But the beginnings were brutal.
"talking" happens only when:
- I truly need to iterate on the output
- There is some frontier-level research (i.e. working on truly novel algorithms/math)
- New UX - even Sonnet5.5 and Fable are still quite schematic and they have difficulty figuring things out from first principles. Astra is smarter but has it's own problems
- Infrastructure problems
If you did please, please share your setup. And how you audited it.
The effect you describe I had at times but with older/dumber models where you could spend a ton of time and the model was either too dumb to do it, or had a model dementia later in the context. Or when I didn’t really understand why it did a bad job and tried to stumble around in the dark - very mentally numbing.
Nowadays the model usually just had some bad tools or information required to do the job.
Also I’ve seen and heard of previously nontechnical people that now build oneoff tools for themselves to speed up their work or coordinate the work better. My fav example is a girl - arts major - who was in customer service, and rebuilt a CRM for a 10-person company. As a cherry on top she created the nicest video tutorial of any software I’ve ever seen :)
But in other cases output was nonprogramming in nature. Since CC can access mail, slides and excel and a bunch of other stuff there is a ton that people can ask it to do and not waste time.
But I’d say only recently CC became decently usable for nontech people - the auto mode plus finally polished browser integration, plus Opus 5.5. Even I was surprised by how well it all works now out of the box (and how cheap - regular Pro license is decently useful now).
Also, it’s been just 9 months of modern claude code, and for many companies less than 6 months since first employees outside of IT got access to Claude Code. What do they expect?
You put in the disc, it plays, if you want subtitles you press the subtitles button, that’s it. I never had it hang or anything like that, and there is no bullshit internet connectivity/ads etc.
The only thing I’d like improved is better streaming/playing local files.
I never encountered any hangs, slowdowns or any other problem really.
Also, it would increase anti AI sentiments, and people’s fears of then using private data - that’s the real cost.
In general you want to be careful about competing with your own clients.
But in many sciences - esp purely theoretical ones - we have infinite datasets that we can generate and there can be no ultimate convergence.
For some benchmarks agents need some sort of a limited internet access - otherwise the benchmarks would be pointless. Good luck implementing a good filter on the internet - if you have a way to do it reliably please share, I'm sure multiple people here would love to hear it.
Here, with a chinese model, it would be even simpler - you finetune a model to activate certain operations when it discovers certain environments, but to behave completely normal otherwise. Kind of like a sleeper agent. And with current tech we wouldn't even be able to spot this.
It is actually faring much better than other tech startups ipo’ing over the last 20 years.
Closer to the surface frameworks tend to add complexity and restrictions, not remove it.
Historically, great defeats caused such cultures to change - look at Germany or Japan. In Russia there was a similar movement in 1991 but it sadly collapsed and they are back at it again.
But because prefill is parallelizable, datacenters will have an edge here for a while. For your one request, a datacenter may push 1.7kW power to a single machine to give you a result in a few seconds. The rest of the time machine will stay busy with other people’s traffic so it’s not a problem. For you, unless you’re a billionaire, you won’t have 1.7kW of processing power running idle 99.9% of time.
In 1-3 years we’ll have datacenters switch to dedicated chips that are way more energy efficient, and such chips will ultimately trickle down to PCs.
BTW - Apple is already introducing LLM specific architecture. They introduced Neural Accelerator (nit to be confused with Engine) that speeds up their GPUs 4x for LLM specific tasks. M5Ultra gets 100 tflops vs nvidia Rtx6000 getting 500tflops. „Just” 5x faster, but it’s beginning to look relevant.
Still 25 slower than a server farm B200. And when you ask Chat/Claude you get the full speed of B200.
What I’m waiting for is consumer dedicated LLM boards, a similar way we had soundcards and first gen 3d accels. No GPU, no 3D, purely LLM compute. I can see that happen in 2-3 years.
A good protip - tell an agent to rewrite the article slightly for better listening flow. Some subtle changes in sentence structure etc can make listening way more pleasant.
Of course they didn’t represent the whole of Russia, but as anyone living next to them will tell you, the further east you go, the more „cooperation through domination” mindset there is. Win-win is a very much western thing.