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

I wonder if it’s related to compacting context after the window is expired?

I blog to think more clearly.

It’s utterly pointless to outsource this to ChatGPT. I need to muddle through with my slow dumb human words trying to explain my thinking against the social pressure of not looking stupid in public. Only then do I truly battle-test an idea.


The main problem isn’t embeddings, in my experience, it’s that “vector search” is the wrong conceptual framework to think about the problem

We need to think about query+content understanding before deciding a sub problem happens to be helped by embeddings. RAG naively looks like a question answering “passage retrieval” problem, when in reality it’s more structured retrieval than we first assume (and LLMs can learn how to use more structured approaches to explore data much better now than in 2022)

https://softwaredoug.com/blog/2025/12/09/rag-users-want-affo...


Love seeing you in these threads! We use “AI Powered Search” as a bible on our team. Thanks for all your contributions to the community.

Thank you. Trey gets the lions share of credit for most of that book :)

For an AI agent to do a good job at customer support, you would need to

1. literally document everything in the product and keep documentation up to date (could be partially automated?)

2. Build good enough search to find those things

3. Be able to troubleshoot / reason / abstract beyond those facts

4. Handle customer information that goes against the assumptions in the core set of facts (ie customers find bugs or don’t understand fundamental concepts about computers)

5. Be prepared to restart the entire conversation when the customer gets frustrated with 1-4 (this is very annoying)


Point 1 (document everything) is the utopia that killed the project. In any complex system, documentation is a lossy compression of reality. The actual truth about how to fix bugs doesn't live in Confluence; it lives in senior heads, Slack chats, and intuition, and AI has no access to this layer of tribal knowledge

Sure but if you attached a hydrogen burner would it be net positive?

Then wouldn’t you have a cleaner energy system then burning the methane directly?


I see Malick in a lot but similarly I see Tarkovsky in a lot of overlapping movies. I don’t think Americans are as attuned to Tarkovskys influence on modern film. I definitely recommend Stalker as an amazing film.

I would also recommend his movie Andrei Rublev, though that is probably even harder to watch because of the length and medieval setting; but, for people that like to see something different, this is very different to current movies.

And from his disciples definitely recommend Zvyagintsev and his The Return and Leviathan.


The final act and ending of Rublev is in my opinion the greatest ending in all of film. So brilliant, you can be thinking the rest of the film is slow and tough to decipher, but it pays off as the ending hits so well.

If you liked Andrei Rublev, you should check out Alexei German's adaptation of the Strugatsky brothers "Hard to Be a God." It's insane, gorgeous, disgusting, teetering on the edge of madness, and monumental.

German's "Khrustalyov, My Car!" is also the purest cinematic distilation of paranoia I have ever seen.


The real takeaway is when a big project can be paused entirely due to one presidents very specific / frivolous whims - we won’t be able to do big projects in the current order. We need a shift in the constitutional order where the whims of one person isnt fused with the bureaucracy

> We need a shift in the constitutional order where the whims of one person isnt fused with the bureaucracy

Correct me if I'm wrong, though there are already protections there. It's just president, senate, congress, SCOTUS all agree on this.

IMHO - most effective constitutional change would be to get rid of first past the post election system, electoral colleges, gerrymandering, etc. I think USA's two party system made it to the place where it is right now, seemingly on the verge of turning into one-party system.


The US has had worse structural power imbalances in the past. It’s gotten over them (after a generation or more) by

- coalition shifts - every election, new groups going and out of each party

- demographic shifts - shrinking / growing / moving around of different groups

- external shock - war / depression / ?? changes incentives of governance (see Civil War, Great Depression)

- hegemons dilemma - the in power party over time goes through in fighting, over confidence, etc (see Republicans becoming corrupt in The Gilded Age)

Alternate constitutional order can mean a lot besides amendments - or even using Federal power. It can be about organizing economic power to reject illegitimacy. It can mean organizing the Democratic Party differently as more of a shadow set of social institutions that support people. It can mean leveraging state power, and building coalitions of blue states. Or other creative approaches to power.


Don't overlook the influence of global soft power and the non-profit sector.

Demographic shifts and a robust social safety net are also naturally reshaping the landscape


The laws exist, SCOTUS majority just doesn’t want to enforce them because their guy is in power.

Maybe but I think it’s more about they think in terms of unitary executive. So if there’s any discretion given the agencies - I don’t know in this case - SCOTUS lets the president decide.

In many ways this is more how a parliamentary democracy exists that a republic.


POTUS power has already extended well beyond even congress.

How? Congress has given tremendous discretion to POTUS. This President is actually using it.

Congress assumed

- it had a legislative veto (any committee could override an agency)

- independent agencies existed.

So it gave broad authority with those assumed checks.

SCOTUS declared legislative veto unconstitutional in 1982. And administrative state is actively going away.

So POTUS can do a lot of damage using the law itself.

This is the new system. Dems need to use it too.


So we had this creeping loss of power to the president over time in the last 20 or 50 years, including investigations in the 1970s or dealing with Nixon. But Congress never really decided this in one big step, it just happened slowly by pushes from the heritage foundation and others. Congress can take back its power. There's a reason why the Republicans are trying to gerrymander the house so the Democrats don't get a majority. It wouldn't just fix it but it would be a start towards starting to block overreach

Tariffs is the obvious example.

Half of the USA, or at least half of its voting population, now supports the idea that the role of government is simply to be an extension of the personality of the Chief Executive. Essentially, whatever Trump feels is the policy of the government and therefor is the law.

I guess you're being downvoted because either: 1) Too many conservative tech bros here or 2) independent voters may not be aligned with this crap yet many voted for him anyway.

Probably both.


What’s sad is how tiny an investment this is relative to their parts of the Federal budget. It will have almost no impact on the Federal deficit (which will be higher than ever this year)

It’s entirely performative


As elections are determined by state and constitutionally mandated, this would be next to impossible.

It’s more likely that legal shenanigans happen like VRA section 2 going away and deepening gerrymandering wars or looking for excuses to disqualify people from voting for “election integrity” and other creative democratic backsliding that would pass muster at SCOTUS.


Without structured outputs, for some classification tasks you can use the hallucinate -> resolve pattern

Step one ask the LLM to classify something from the prompt “creatively”. For example, ask it to classify the color or category of a product in an e-commerce catalog or user request. Give examples of what valid instance of these entities look like, ask for output that looks like these examples (encourage the LLM to engage in creative hallucination). Often helps to get the LLM to pick more than one and for it to choose many different, realistic diverse labels.

Step two, with hallucinated entities, lookup via embedding similarity to find the most similar “real” entities. Then return these.

It can save you a lot of tokens (you don’t have to enumerate every legal value). And you can get by with a tiny model.


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

Search: