HNHacker News
TopNewBestAskShowJobs

inciampati

1,697 karma · joined June 29, 2017

submissionscomments
inciampati··on Dow plunges 2,200 points, Nasdaq enters bear market
The "intentionally tank the economy to roll over national debt at lower rates" theory completely misses how government bonds actually work. A bond is essentially a promise: you lend the government money for a specific period, and they pay you back with interest at the end. When that bond matures, it's simply paid off. There's no refinancing happening.

What people call "rolling over" debt just means the government is issuing brand new bonds to raise money to pay off the ones that are maturing. These are completely separate transactions, often with entirely different investors. The government continuously sells new bonds and pays off maturing ones as part of normal operations.

Creating a recession to somehow game this system would be spectacularly counterproductive. Tax revenue would collapse while unemployment benefits and other social spending would skyrocket, creating even larger deficits. Any minor interest savings would be dwarfed by the economic damage. Not to mention the Fed sets rates based on economic conditions, not to help government borrowing. This isn't some clever financial strategy—It's just bad economics built on a fundamental misunderstanding of how sovereign debt markets function.

inciampati··on Do viruses trigger Alzheimer's?
The problem is simply observational. We don't even have reliable DNA and RNA sequencing of our own bodies. And we cannot reliably observe things in a host without knowing, to some extent, what we're looking for first. Even that space is so large, it's very hard to ascertain accurately. Biology is always suffering for lack of clear observations.

Also, adding complexity is the difficulty or even literal impossibility of observing the direct interactions of elements of the system, which operate at a quantum scale, that you would disturb and do disturb when attempting to observe.

inciampati··on Tencent's 'Hunyuan-T1'–The First Mamba-Powered Ultra-Large Model
What's the best technique to train the model to stop responding? A bit of fine tuning on texts with EOS markers?
inciampati··on Deep Learning Is Not So Mysterious or Different
An example, which is interesting, in which "deep" networks are necessary, is discussed in this fascinating and popular recent paper on RNNs [1]. Despite the fact that the minGRU and minLSTM models they propose don't explicitly model ordered state dependency, they can learn them as long as they are deep enough (deep >= 3):

> Instead of explicitly modelling dependencies on previous states to capture long-range dependencies, these kinds of recurrent models can learn them by stacking multiple layers.

[1] https://arxiv.org/abs/2410.01201

inciampati··on Big LLMs weights are a piece of history
And I'd love to see data compression terminology get an overhaul. Do we need big LLMs or just succinct data structures? Or maybe "compact" would be good enough? (Yeah LLMs are cool but why not just, you know, losslessly compress the actual data in a way that lets us query its content?)
inciampati··on Backyard Cyanide
I recently had a friend-of-a-friend's apple seed liquor. It smelled of almonds...
inciampati··on I've been using Claude Code for a couple of days
I'm impressed by how many people who are working with Claude Code seem to have never heard of its open source inspiration, aider: https://aider.chat/

It's exactly what Yegge describes. It runs in the terminal, offering a retro vision of the most futuristic thing you can possibly be doing today. But that's been true since the command line was born.

But it's more than Claude Code, in that it's backend LLM agnostic. Although sonnet 3.7 with thinking _is_ killing the competition, you're not limited to it, and switching to another model or API provider is trivial, and something you might do many times a day even in your pursuit of code that vibes.

I've been a vim and emacs person. But now I'm neither. I never open an editor except in total desperation. I'm an aider person now. Just as overwhelmingly addicted to it as Yegge is to Claude Code. May the future tooling we use be open source and ever more powerful.

Another major benefit of aider is its deep integration with git. It's not "LLM plus coding" it's really like an interface between any LLM (or LLMs), git, and your computer. When things go wrong, I git branch, reset to a working commit, and continue exploring.

note: cross-posted from the other (flagged) thread on this

inciampati··on Another late-night Claude Code post
Aider does scan your entire codebase. What did you try to do that didn't work?

A typical workflow for me involves developing a plan document using copy/paste from the command line into a (cheap/fixed cost) webui and discussing it with the AI. Then, the result gets pasted back into aider to implement the changes in code. Aider, for me, mostly replaces the text editor, and the mindless process of squashing compilation and test bugs.

inciampati··on Another late-night Claude Code post
I'm impressed by how many people who are working with Claude Code seem to have never heard of its open source inspiration, aider: https://aider.chat/

It's exactly what Yegge describes. It runs in the terminal, offering a retro vision of the most futuristic thing you can possibly be doing today. But that's been true since the command line was born.

But it's more than Claude Code, in that it's backend LLM agnostic. Although sonnet 3.7 with thinking _is_ killing the competition, you're not limited to it, and switching to another model or API provider is trivial, and something you might do many times a day even in your pursuit of code that vibes.

I've been a vim and emacs person. But now I'm neither. I never open an editor except in total desperation. I'm an aider person now. Just as overwhelmingly addicted to it as Yegge is to Claude Code. May the future tooling we use be open source and ever more powerful.

edit: Another major benefit of aider is its deep integration with git. It's not "LLM plus coding" it's really like an interface between any LLM (or LLMs), git, and your computer. When things go wrong, I git branch, reset to a working commit, and continue exploring.

inciampati··on Substance in broccoli, celery inhibits graying of black hair
IIRC hair dying is not great during pregnancy. It is slightly, but measurably oncogenic in children of mothers who died hair during pregnancy and breastfeeding. https://pmc.ncbi.nlm.nih.gov/articles/PMC3543291/
inciampati··on NIH insiders reveal process stalling grants
Genetic diversity is a scientifically measurable feature of an organism or population, often diversity is stylized as π. I've heard of grants that mention this diversity in terms of mouse populations which are currently frozen. The solution is to talk about 1-π, or "purity". Scientifically the same. Politically acceptable. lol
inciampati··on Ways to annoy your senior engineer
It seems that people in these kinds of positions feel very powerless, frustrated, and cranky. And they love to vote for posts that say these kinds of things. They validate the frustration that they experience working in these enterprise environments.

In theory, you're asking for hacker news, it just happens to be that there's many different communities represented here.

The tyranny of the front page means you're probably missing a lot of the posts that you might want to see. And people aren't making many of those posts because they know they won't get seen by many people.

inciampati··on The Generative AI Con
by giving the models durable memory, they become "agentic". in any case, they don't make as much of a mess when their output is getting written into git.
inciampati··on 670nm red light exposure improved aged mitochondrial function, colour vision
Weirdly they didn't test nighttime exposure as an intervention. OK, "weird" means probably it was costly/hard to get people into a lab setting at 9pm... But it does make you wonder!
inciampati··on DOGE Has Started Gutting a Key US Technology Agency
To date there is no evidence that any court order has been respected. The crisis continues.
inciampati··on Teen on Musk's DOGE team graduated from 'The Com'
It's thanks to this kind of guidance that HN survives as a focused technical hivemind.

At the same time, issues of this kind of revolutionary scope are important for users to process. We can learn a lot from each other.

Irrespective of politics, it's necessary to hedge systemic risk that's appearing due to destabilization of the US. That affects so many of us that it's hard to ignore.

Keeping some persistent outlet (front page post) for discussion of this major topic is important to give people a politically agnostic and technically proficient space to integrate what's happening.

Thank you for filtering the noise and fear with the posts.

inciampati··on OpenAI's o1 and o3 models forget what they thought
Reasoning models from OpenAI forget what their past reasoning. The reasoning tokens are removed from context after the user-visible response is generated. It's not what I expected, given that memory of reasoning is the key to chain or thought's success.
inciampati··on OpenAI O3-Mini
LLMs are actually the new computers. Compilation is only one program they can run.
inciampati··on OpenAI O3-Mini
o-whatever are doing the same thing as any LLM, it's merely that they've been tuned into using a chain of thought to break out of their complexity class (from pattern matching TC0 to pseudo-UTM). But any foundation model with a bit of instruction tuning is going to be able to do this.
inciampati··on OpenAI O3-Mini
LLMs (transformers) literally cannot balance parentheses. That's outside of their complexity class (TC0). You'd want a real UTM to count parentheses!
inciampati··on OpenAI O3-Mini
> Is this a case of PEBKAC?

It's 2025 and every useful conversation with an LLM ends with context exhaustion. There are those who argue this is a feature and not a bug. Or that the context lengths we have are enough. I think they lack imagination. True general intelligence lies on the other side of infinite context length. Memory makes computation universal, remember? http://thinks.lol/2025/01/memory-makes-computation-universal...

inciampati··on OpenAI O3-Mini
This has been said by every long term user of these sites. And not at the same time. It was always better in the past. It's probably partly true... yeah, quality can decrease as things get better. But it's also partly an illusion of aging in a changing world. Ten years is long enough to completely change the way we write and express ourselves.
inciampati··on OpenAI O3-Mini
This is so good! Finally the word I need to describe what I'm living. And yes I'm copy pasting articles into LLMs for summary.
inciampati··on Memory makes computation universal, remember?
Well, part of the reason that this writeup doesn't mention the brain is that the author is ignorant of precise details of how the brain works, and chose not to speculate :) Which is exactly why these papers are such a useful reading list. Cheers!

It's fascinating that we have had so many attempts to add memories to NNs, but it was the end-to-end-learned one based on textual context that eventually drove us to models that have emergent reasoning abilities.

RNNs basically have a memory, in the form of their hidden state, but it's been hard to train them efficiently in parallel. There may be a better way though, if we can.

inciampati··on Memory makes computation universal, remember?
I wrote this after spending December thinking about why o3's expensive memory usage matters. The key insight feels trivial (CS101: universal computation needs stable state + memory), but watching xAI (and now Stargate) build gigawatt facilities while o1 explicitly discards its own reasoning made me realize we've forgotten something fundamental... https://arxiv.org/abs/2412.17794
inciampati··on Memory makes computation universal, remember?
Wonderful! Thanks, I'll definitely read the Maturana and Valera work, would be good to work through this line of thinking.
inciampati··on O1 isn't a chat model (and that's the point)
https://platform.openai.com/docs/guides/reasoning/advice-on-... this explains the bug. o1 can't see its past thinking. This would seem to limit the expressivity of the chain of thought. Maybe within one step it's UTM, but with the loss of memory, extra steps will be needed to make sure the right information is passed forward. The model is likely to start to forget key ideas which it had that it didn't write down in the output. This will tend to make it drift and start to focus more on its final statements and less (or not at all) on some of the things which led it to them.
inciampati··on O1 isn't a chat model (and that's the point)
I found this documentation from openai that supports my hunch: https://platform.openai.com/docs/guides/reasoning/advice-on-...

The reasoning tokens from each step are lost. And there is no indication that they are different tokens than regular tokens.

inciampati··on O1 isn't a chat model (and that's the point)
Yes it is, because the post discussed this approach precisely because unrolling the actual chain of thought in interactive chat does not work.

And it's doubly relevant because chain of thought let's transformers break out of TCO complexity and be UTM. This matters because TC0 is pattern matching while UTM is general intelligence. Forgetting what the model thought breaks this and (ironically) probably forces the model back into one-shot pattern matching. https://arxiv.org/abs/2310.07923

inciampati··on O1 isn't a chat model (and that's the point)
Where is that documented? Fwiw, interactive use suggests they are not available to later invocations of the model. Any evidence this isn't the case?
← PreviousPage 3 of 20Next →