Hacker Newsnew | past | comments | ask | show | jobs | submit | spockz's commentslogin

Find a way to store it properly or convert it into higher quality things. It’s not being done now because it is easier and cheaper to just dump it into ships.

For the finished jre yes. But one would hope these kind of checks are also run on a hand rolled decoder function. Fuzzing that one and finding logic bugs is a lot easier

The second hand market is flush with M1s and M2s. Even refurbished from Apple itself. I wouldn’t worry too much about it for the near future.

And I’m not sure prices will come down to pre AI level for a long time. The vendors are too comfortable with the current pricing.

Also note how quick prices at the pump go up when there is a moment of oil crisis and they take weeks to go down after things resolve.


Not where I live. Here even Apple doesn't sell refurbs and whatever is available from this party is worse than a toin-coss, if not downright shady (which is most cases) – an with zero warranties (literally).

And yeah, couldn't agree more with prices never going down. In our country oil prices have been like Hotel California: you can go up any time you like, but you can never come down.


Oil prices, as they say, tend to "rise like a rocket and fall like a feather"

Ah the irony. (AI being trained illegaly on copyrighted content.)

The other one being naming and caching.

On pro here and I can do about 2 design tasks a week before I hit usage limits with Argon. The usage limit with carbon is probably gone just loading the context.

I've been using Argon (Medium) pretty heavily the past couple of weeks to create a website and an app and am not coming close to my Ultra limits. I've used High for a couple of code reviews and security audits.

I'm not sure if I'm a sucker for paying $100/mo for this, and Google's harnesses are so underwhelming compared to Claude Code, but I'm not having a bad experience. Compared to the beginning of the year when Antigravity was almost unusable and Gemini Pro 3.1 was noticeably behind other options.


> Google's harnesses are so underwhelming compared to Claude Code,

Could you please elaborate? A heavy user of Antigravity myself I had the opposite impression. Quite possibly I only scratched the surface of what Claude has to offer. For example how would you compare `/remote-control`? Would you say you prefer the Claude version ?


Have you tried running antigravity without?

  agy --dangerously-skip-permissions
Claude Code auto-mode works fine, no need for yolo mode.

Yes I run Antigravity without. Note that AGY auto mode is disabled by default. If you find it asks unreasonably often that's most likely the issue.

(It could also be that the classifier has a higher rate of false positives. Please let me know if that's what you meant. This tricky to reason about though, because eg users won't typically complain about the false negatives.)


Argon is slow but surprisingly my progress is noticeably faster. If you aren’t prompting extra ambitiously with top tier models, you should probably just use deep seek.

Not quite at the build me a whole app level but you need to definitely zoom out a few times and prompt from either a feature set or many overlapping details at once.


Yep, I'm on Pro as well, and Argon been hugely useful, but I get about 10 prompts with it a week. And that means that I have to use 3.8 flash for most of what I'm doing, which is obviously not a great experience for any kind of long horizon tasks.

I’ve been trying to get something similar to work with local markdowns instead. Getting them to write something is easy. But I cannot reliably trigger them to write when they make “sufficiently important” decisions. There is little self reflection (or actual intelligence).

How do you get them to trigger writing to the log in the right moment and manner?


The karpathy wiki is really just local markdown at its core. The semantic search qmd provides is really the fancy bit.

My system is fairly elaborate, but the simple trick is to build the loop so they have mechanical checks that prevent anything from counting as done until it meets standards. There are many ways to do that. A lot of folks just hook specific tool calls and make it output relèvent instructions.

In my loop I use an Orchestrator and a Builder sub agent. The Orchestrator assigns the Builder work from the spec one phase at a time. The Builder has to build each phase with red then green unit tests as specified by the Orchestrator. The Builder sub agent only knows that one task. It can't consider the task finished until gatekeeper passes.

Gatekeeper is a simple python script that the Orchestrator is allowed to update whenever it sees an anti-pattern. It validates that all the unit tests pass, that it HAS sufficient test coverage, that it lints properly (including custom rules the Orchestrator can add for repeated failure), that the docs exist for that phase and are long enough, and finally does some ast grep searches for common goofs, etc.

Sometimes the builder will have to try many times before it passes, but each time it fails gatekeeper outputs an instruction telling them which rule they violated and the full requirements.

Eventually gatekeeper will pass and the Orchestrator does a final review before assigning the next phase to a new Builder. The most common failure at this point is poor unit tests that don't test for the right thing. The Orchestrator has explicit instructions to check for that and does catch it most of the time.


Hi spockz, I separate two thing, logging permanently that agent do. And write a decisions or important point, as you say is not easy because «important» are subjective. In my side some important point are sometime useless but not wrong. Forcing to record proof can help.

Daybreak blue (for defensive action) was for a long time a specific model selection, basically Sol 5.6. Later it became a toggle available for any 5.6 model. Last time I checked it wasn’t available for 6+.

Daybreak red for offensive action is a better trained model of Sol 5.6 for offensive action.


I see. Interesting.

Yes you can. You start a server with codex cli. Do a round trip with an authorisation token and you are in.

I couldn't manage to get it working with my vps and the android chatgpt app. Still using ssh and more comfortably pi remote with port forwarding.

What does that mean? That it only reads part of it into context or RAM?

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: