ghm2199

Structurally speaking we learn nothing like AI, we don't use vast amounts of information to pick up completely new skills. We also make decisions by using prior knowledge and emotions.The latter part is important, Thinking fast and slow cannot operate in a world of AIs as they stand today unless we are willing to grant them rights — because you have to teach them to make decisions based on all kinds of emotions — which is tricky at best.

> The tool call is in the text

Noob question: do good harnesses automatically optimize for bad tool calling behavior automatically?

Google did make location history private in the maps app(delete on server side) at some point, if I recall correctly everyone can at the least opt in(i may be wrong and it may even be opt out!). I believe out of the exact concerns of warrantless sweep of all location data of all users around a lat long for server side storage by law enforcement.

It went under the radar but is a good example of _some_ incremental positive improvement over centralized storage.

Altman, Amodei, Pinchai etc can get together and just put a self imposed ban on themselves and ask the us government to sanction any entity/state attempting to use/release a Recursive Self improving model until the risks are well understood.

This does not mean they can't keep developing it. But it would mean they do not release them to the general public or the public domain. The US government can continue to use and fund these models that are more powerful and use them in situations they understand and a restricted setting. The pentagon has a larger budget than any VC.

Remove the profit motive on self improving recursive models gained from retail/commercial use!

I use pi+codex and OpenAI codex models do much less of this in my experience.

One thing i am curious about is if anyone who is fluent in a romance language(en/es/fr etc) but their native language was a non-romance one(e.g. japanese or hindi) felt it better using the latter exclusively?

Sure, but YC can enforce it if they want to. VC after all is a private body governed by private individuals, they can make moral choices. VCs could even support this by offering terms to offset the costs. I cannot believe that a company cannot be profitable but at the same time not build for at least some better privacy guarantees.

Are you guys feeling all high and mighty doing all these predictions? I think the money they are spending on trying to inform policy could be better spent on getting all major labs together with the government, and making sure that

1. At least 10% of your net spend is on safety and harm mitigation before you accelerate yourselves into doing real harm to real people tomorrow.

2. Pause all AI enhancement until a proper mechanistic study can be done by a working group funded by you all to understand how AI model Safety can be enforced by something less lame than a Constitution/Spec.

And also the fact that agents are solving anubis level 5 challenges Tailwind would not be able to figure out how much traffic is from people vs bots. So that would mean the traffic should be up...

Noob question, I recall https://news.ycombinator.com/item?id=49491791 mentioned how the kernel source code and diffs were being crawled excessively. All else being equal, won't all the AI crawling agents mean traffic to docs should at least be stable(presuming changes happen all the time)?

Perhaps its time that YC mentioned explicitly that if they are going to fund a company which holds specific PII data(i would include precise location history to start with) they would have to guarantee:

1. An erase-me feature if they gained over say some X unique/active users.

2. Disclose to their best legal abilities on warrants served in the countries that allow such disclosure. Allow for strong whistle blower protection if warrantless wiretapping becomes a thing.

Congrats on the launch. I had a vague idea about personalized medication, did not know it was the VC style scaling of 10m$->100m$ ARR business!

There are some messages here complaining about the website being claude(ish). I know HN has people that care about this. They should. Mostly they would be right. But i don't see it being relevant here. Complaining about this business's website is like fluid-mechanics engineers at NASA complaining about friction dynamics on office floor-carpeting instead of rockets.

awesome!

Could it put small to medium sized packages in a package room of a condo?

Hopefully upto 24" in any dimension and 5-10 lbs(and ignore the others it can't handle)?

The set up of the small room: three stacked total height 5' rack shelves on both sides and a 2' wide passage in the middle.

Not saying off the shelf but how hard would it be to teach it how to do that?

Why can't we just be satisfied with wayy old mac versions in VMs on linux, ones where you played prince or persia or oregon trail, the fun looking black and white versions and some later ones their steel windows UIs, the theming, the file explorer, the note taker app.

Isn't that level of fun one gets from looking at that enough?

Thank you google for hosting the downloader that exists for making it easy to use for automated testing('cause all the other people still use it). Its the only use case that exists for me(actually my ai) to use it.

The reason why people(or ai agents) have not been able to systematically destroy bad research in fields where fraud is so rampant is because they data are not published or easily accessible without emailing the author and asking them.

I sincerely hope that any self respectful journal that publishes papers should have a minimum bar: submit the data on a public link with no paywall or blockers and feature the link prominently on your paper.

I think for voice-ai every one builds their own harness, and thats why there isn't a common one out there. Also if you are a serious company that has a llm in your production loop, you would have to build an in house thing because its so critical to your product. Its like performance-engineering, most performance engineer work is done in house and it varies wildly. Another reason is that because voice quality is a vibe measure(intonition, pitch, human variation), its impossible to make the whole thing deterministic.

I use pi and built a harness for just an llm that calls a bunch of tools. That got me 50% of the way and it would be fast. Then build it for STT and TTS, this will be slower but it will get you far. There are a bunch of tools out there for building basic harnesses.

Realtime + Voice AI usecases is where latency is most important. I use Handy on my desktop and i can tolerate a latency of a few seconds every now and then. Your P99 should on TTFB should be really low to compete for voice ai realtime

One thing that evals are super important from the get go are where the harness+model inference is part of the product, e.g. if you are doing voice ai, building out a test harness to test the system is a non trivial first step.

If you switch from claude to codex(or really any other frontier lab) after a spending a little bit of time with how claude writes, you will immediately smell claude's writing from a mile away.

Technically does this not create a problem for their EPS because of the poor free cash flow? Current market consensus is $81B in Net Income for the year but a -$6.7B cashflow for 2026[1]. But the stock is up 1%. So, is the -$12B is priced in(somehow)?

[1](https://www.marketscreener.com/quote/stock/META-PLATFORMS-IN...)

I think the posted article is also the description of America's labor market. Many people don't use it on a day today basis and most people cannot engage their system 2 to assess risks over a few weeks out, let alone 5-10 years.

But I found This is a real practical everyday example of why statistical uncertainty is important to know about. It teaches you how to compare — risk adjust — two vastly different investments e.g. BTC vs S&P or indexing vs value strategies. Then you sleep at night.

It is also not intuitive and many people are very anxious and FOMO driven when it comes to money. So you need to internalize the idea for it to sink.

The drawdowns are a statistical measure and you could be unlucky to catch a big depression style thing once in 30 years. And T is typically longer than 5 years, typically 10 or more.

One of the simplest ways to measure risk in stocks, bonds, crypto, mfs etc for ordinary people is Maximum Drawdown. Once you understand how it works, it takes the stress out of making and managing your own portfolio P — espexially your non retirement account. The question you have to ask your self is: "Given some portfolio with returns of X%, am I ok with this asset being down by Z% over T years — i.e the _computed/inferred_ drawdown?" If the answer on one end is no I cannot afford P to have any drawdown at all, then just put your money in a cash/MMF and call it a day. Generally people are ok with some risk on some percent of P and stash the rest in cash, and you _risk adjust_ for Z and T. The portfolio choices are surprisingly simple.

No other question matters. There are some risks but the key element is the understanding of the statistical variance because that allows me to say that "ok 50% of P can afford to be down for 3 years and I won't be homeless".

Most people don't know this but most money managers(managing money for ordinary Americans) that you hire compute this number _once_ and make tiny adjustments to your portfolio(I am talking once a year maybe) and take 1-2%.

I follow this guy called Dave Stein who has a B2C product called money for the rest of us that taught me this. I am not affiliated with them in any way.

Is it just a matter of incentives by the government in indonesia? Like is it that bomb fishing in indonesia is incentivized because of, for example, profit margins in fishing vs say tourism?

I know you mentioned Indonesia never got the yellow card it should have and that certainly creates less incentives for them to straighten their act. But I am trying to understand if the government is still largely to blame or not?

Fair enough.

The collective harm that this kind of language does to a human skill of writing by oneself is probably immeasurable (that is until some clever social science researcher finds a controlled way to quantifying this).

The full statement:

> This isn't just a new feature — it’s a new way to build software: open, collaborative, and powered by both human ingenuity and judgement and agent scale. Teams that build this way won't just move faster. They'll build things no one else can.

I am not sure how just drawing contrasts between two things without actually drawing any kind of causal relationship to explain _why_ something is better or _how_ it does it, came to be a good thing. It is the kind of vaccuous statement that some poor tired sod with his remaining system 1 capacity just YOLOd into the blogoshpere and, like you said, the weary who don't know any better get FOMOed by. PSA: you arent missing anything.

Interesting. I do feel letting a machine generating notes is taking the joy out of improvisation, is it not?

I think one of the great, early, joys of learning a piano is gaining the following intuitions: The seemingly harder path of learning sheet is actually faster. Your mind _should_ learn to think in two dimensions Spatial — where fingers go — and Time — pitch and tempo – ** when learning. The _internalization_ of Space and Time queues guide the fingers in a dance that is vastly satisfying. This skill leads you to the final part of the journey that is improvisation and the one more exciting than what i am on now.

---

**

Space: Your finger placement on keys right, e.g. knowing how to go from landmark/anchor notes(mid-c, G, F etc) and then go to the others above and below it. Crudely this is some what like typing from your landmark f and j qwerty keyboard

Time: The out singing/verbalizing of the notes/beats on a time measure as you play them(per the time measure). e.g. you can say out loud 1-2-3-4 for 4/4 measure, if the measure has quarter notes say out loud. And `1-e-and-a-2-e-and-a-3-e-and-a-4-e-and-a` for a 4/4 with 1/16th note granularity. Do this as you play the notes and you get a sense of tempo.

Goal 6.4 reminds me always of the numerous times AI generated n PR's for a feature and I revolted and threw my laptop because it was incomprehensible or unworkable when viewed as a process/workflow.

Also the removal latency is on a log scale. Which is quite insane.

Wow! 4GB for 10 million documents. This means one could build a reverse index much faster than before and devx processes like debugging, performance testing would become much smoother. Can't wait for the sqlite bindings to come out!