I know it's incredibly presumptuous for me, a nobody, to say this to 25 Fields Medalists, but:
Perhaps you are misaligned.
Who decided the goal of math must be human insight?
First off, some mathematical truths might simply be far beyond our biological comprehension.
Second, for us non-mathematicians, the value of math isn't in understanding exactly why a result is true, it’s in how those results can be applied to actually improve our lives.
Isn't this why we have math in the first place? To solve our real problems? Over time it morphed into this pursuit of pure theoretical insight, probably out of necessity at the time, but is it still necessary?
> Second, for us non-mathematicians, the value of math isn't in understanding exactly why a result is true, it’s in how those results can be applied to actually improve our lives.
This is just the problem with the Navier-Stokes solution. We now (supposedly) know the truthiness of the statement, but we have no more understanding of why or how it applies to our quality of life. This is exactly the issue Tao is pointing at.
There’s one thing I constantly see agents tripping over, I’m not sure what the right word for it would be, but it basically boils down to “making changes in the right places”. They seem to have very poor grasp of where things are supposed to be and they have a tendency to work against the existing architecture. Even in a world where agents are the only ones touching the code you can see how this ends poorly. Unlike correctness I’m not sure there’s an easy way to verify.
I tried writing a few skills to encourage agents to spend time thinking about this but it doesn’t seem to generalize very well.
I’m not a fan of OAI to say the least, but having worked at similar companies, my guess is that it’s just too difficult to prove/disprove beyond a doubt, and they have other priorities
I think the line between regular LLM "endpoints" and agents/harnesses is going to become more and more blurry until it's a meaningless distinction.
When you're using ChatGPT/Claude/Gemini etc. you're basically already interacting with some backend harness with tools etc., not a raw LLM. Just give it a computer and be done with it.
I already find myself using Claude Code / Antigravity (via web) instead of Claude / Gemini, even for tasks unrelated to coding. Why use a limited version?
Because in most of those API, even many implementations of the Responses API, you lose a lot of control of where your data is going. e.g. an Agent or the Responses API may automatically invoke a tool call that leaks your data to an external service on the internet, without having an option to intervene.
If you want to have control over your data, you have to have control over your harness.
This bothers me so much with the existing offerings. I start with the chat interface then as soon as I want to get technical/run scripts/automation, I have to copy the context into a fresh code session. So cumbersome.
In my experience, Claude Code is great at making a first-pass at a project, but once you start asking it to make changes, it explodes. A bug fix that should only be 2 lines turns into adding 3 functions totaling 100 lines.
Something as simple as "make the button blue" should be done manually.
Surely you still have to give feedback to Claude. How do you do that without talking to Claude directly? By using a different model? But wouldn't that AI have no more common sense than Claude?
Initial dev required talking to Claude. Now I add a ticket in the board, it makes me a mockup/writeup, I approve, and it gets me a temporary webserver, iOS/Android build, etc. to verify it.
Review loops, agents that enforce my pet peeves and testing/debugging processes, etc. all run automatically... and then Codex strips down the prose at the end. There's not zero AI generated output, but it's already been critiqued and verified by a whole cluster of independent actors before it gets to me. When I have feedback, I file a ticket.
I wanted to get out of the "what the fuck, why?!" loop. Now I let the agents handle that.
I’m of the feeling that talking directly to these things all day is a path to madness.
They talk to each other, come to consensus, and give me a structured thing to look at that has already been vetted and tested and screenshot evidence and a verification plan. It’s not conversational for me.
Ok, but my problem is that what it builds is not right. It's never right on the first try, or the second or third tries. There's always a ton of incremental corrective feedback needed.
Do you run into this problem? If not, then I need to figure out how to stop running into it myself.
If so, then how do you give it all the incremental corrective feedback without a conversation?
People have been sleeping on Gemini lately but these last few Flash releases (which were very rapid) are damn good.
These sort of fast and cheap models are great for tasks that are verifiable and can be retried infinitely (like coding), you can basically get frontier results with a good harness (at a fraction of the time and money).
I highly recommend just getting out of Anthropic's (or anyone's) vendor lock-in. Use opencode or pi. You can still use your subscription pricing using a proxy. I switched to opencode and haven't looked back.
I use OpenCode with my ChatGPT subscription, which is officially supported on both sides and a damn good deal (OpenCode Go is great too).
OpenCode has “providers” for many (many!) other services, but these are almost all unofficial and against ToS (Anthropic being famous for ban-hammering people).
I've curated my auto-approve list to specific commands by approving them with "always allow in this project" (I never want it e.g. committing/pushing to GitHub, removing files, etc without being in the loop) but Antigravity does have both a standard "auto-approve" mode _and_ a "Turbo mode" which disables ALL approvals of all kinds.
I resisted at first but now I main antigravity for work. All the software (web, react-native app, client cms platform, postgres backed, multiple ETL systems, a few chat/websocket backends) for my company is loaded into a single project. I spend time writing prompts and forming plans back and forth with the agents then I click GO. Over the past year we've gone from it taking minutes and needing a reasonable amount of back and forth and fixing, to it taking 10-20 seconds and outputting near perfect work within my system accross app/service boundaries.
Makes it easy to switch between models and I like it for exactly the reason that you're saying - I prepay and so can't accidentally spend my food budget.
Aside from the other reply, pi makes it _really easy_ to build your own usage tracking and limit machinery.
I would normally advise against such efforts for a variety of reasons (such as inaccurate tracking, etc), but specifically under pi, this mechanism has been extremely well behaved and accurate for me.
If you're on their subscription plan - agy cli or antigravity ui is the only choice i think.
Anyway - if you're a dev - you would be writing your own agentic env right ? that's the best way forward. I wont tell you more than this . but if you're not - you are losing out .
I agree, I use my own harness (link to the most stable version, from my Racket book: https://github.com/mark-watson/Racket-AI-book/tree/main/sour...) and except for handling user interrupts correctly, writing harnesses that are customized to your workflow is fairly easy.
You can use any model with Claude Code. Most chinese one have a Anthropic compatible endpoint and for Google and OpenAI's models you can get a compatible endpoint with a proxy like Bifrost. No need to change your harness.
The vscode chat is great. Not sure if it's also called copilot, but you can plug in any models there and they get sandboxed, tools and link to your code. Great stuff.
Antigravity has been also rapidly improving lately, and your can also use any of the open coding harnesses. But I mostly meant “harness” as in your workflow/loop setup.
Update: After some backlash, Google has clarified that they will only ban your "Antigravity and/or Gemini CLI accounts," not your Google account. How very generous! Keep being tone-deaf then...
I might be wrong about this, but obviously Google would like to provide inferencing at the lowest cost to themselves, so perhaps their slow ‘pro’ releases and rapid ‘flash’ releases is an attempt to guide people to use more profitable models?
reply