HNHacker News
TopNewBestAskShowJobs

perrygeo

3,193 karma · joined April 28, 2014

submissionscomments
perrygeo··on GNU Radio in the browser
I've tinkered with SDR some and did a bit professionally. I always found GNU Radio to fall into the uncanny valley - too GUI to scratch my programming itch but too low-level to be useful for exploration. Between libraries like `pyrtlsdr` and desktop apps like `sdrangel`, I can't seem to find a niche for GNU Radio.
perrygeo··on Pre-Release of Polars 2.0
Empirically false. Python's type system is not sufficient to describe runtime variants, period. I have worked on at least three system that used mypy in strict mode yet hit runtime errors constantly.

I get where you're coming from. A good type system should provide invariant guarantees about runtime behavior. But Python's does not.

perrygeo··on “Next-token predictor” is the wrong mental model for LLMs
"LLMs are next-token predictors" is a perfectly accurate mental model. But that doesn't preclude higher-level models such as "LLMs emulate artificial general intelligence". Both can be true.

In systems, we can have facts which emerge from other facts at different levels of abstraction. The causal relationship is not linear. It's not entirely clear that next-token prediction should result in anything close to "intelligence". Yet it does.

Life is another good example. Some might say "biology is just organic chemistry" while others might say "biology is an interconnected planetary system which captures low entropy energy". Both are true.

As a result of emergent phenomenon, we have to take the stance of explanatory pluralism; using the explanation that works best in context. There is no single mental model that works everywhere.

I will continue to think of LLMs as next-token predictors because it's (sometimes) useful, and empirically true. But I also think of them as "pattern matchers", searching for language patterns and trying to replicate them. This is also (sometimes) useful and empirically true. There's likely an infinite number of mental models; our job is to pick one that's both true and useful.

perrygeo··on The asteroid currently hitting front end web development
Or another angle - will we even need a framework? Frameworks exist to give humans better a better interface for common patterns. But LLMs can find and replicate patterns all day in vanilla Javascript without getting bored or annoyed by bad ergonomics.
perrygeo··on Pre-Release of Polars 2.0
> I rarely have type problems in my programs that python typecheckers can't catch.

> I do have type and dirty data problems of the dataframes/csvs/parquet I'm reading.

Something about those two statements makes my brain explode. You rarely have type problems, except when dealing with data? Isn't the whole point of a program to deal with data?

perrygeo··on Rich Dad Poor Dad author shocks fans with $1.2B debt
What's ironic is that the book goes to great lengths to explain this. The entire point of the book is: use leverage to grow assets.

The alternative, not taking debt and using income to directly pay expenses, is literally called out as the "cash flow pattern of a poor person"!

Agree with his financial strategy or not, it's very clear that accumulating debt was part of the plan, not some failing of the plan.

perrygeo··on Pre-Release of Polars 2.0
Polars is much better than pandas in this regard. But using it via Python still carries the risk of runtime errors.

Python's linters leave many holes unfilled - it catches some stuff but still leaves everything else to the runtime. Literally every Python application that I've worked on in the last 3 years a) is 100% type checked yet b) still hits massive numbers of novel runtime errors on the regular. Python's anemic type system is in no way a substitute for an advanced compiler.

perrygeo··on Pre-Release of Polars 2.0
I agree. But the data science/engineering space is enamored with Python. Makes good sense when doing interactive work. Makes no sense in production. But by that point, most developers consider it a sunk cost and just keep their Python apps limping along rather than rewriting.

That's why polars is a great option. Start prototyping in python, then a relatively easy port to a Rust app when the need hits.

perrygeo··on Pre-Release of Polars 2.0
For me, the superpower of polars is production stability.

Pandas tends to push all problems to runtime, with all sorts of hidden heuristics. Particularly around column types and missing values. It's very hard to know if you've tested all the edge cases. The only way to test your code is to throw all variations of data at it. Fine if you're sitting at a notebook and have the patience to validate and "clean" the data on its behalf. Not so fine if you get paged at 3am because your data pipeline failed when it expected an int column but got float.

Polars is more strict by default and front-loads costs through its planner. The resulting apps are noticeably more stable in production. You can test code and reasonable assurance that it will work on data in the wild.

I don't really have any interest in the API ergonomics or syntax - both are fine. It's all about how they deal with data variation at runtime. Can you write general code that doesn't break on variants? Pandas, not a chance. Polars, absolutely!

Bonus round: polars has a Rust API too, the compiler can effectively prove that your program handles every edge case. It's common to write rust polars apps that run unattended for years.

perrygeo··on Coordination Headwind: How Organizations Are Like Slime Molds
> The military can afford to learn. You can't do that in a business.

are you literally saying that businesses cannot afford to learn?

Businesses must afford to learn. Under capitalism, that's the job description. Those that fail to learn... just fail.

The entire reason for the existence of a market is to provide information. Buyers and consumers exchange information, coming to an agreed market price, if you're out of the loop, sorry. This was capitalism in 1790. This is still capitalism in 2026. Not much has changed in that regard.

perrygeo··on Aphantasia Beginner's Guide
That's just "counting". But for many people (the vast majority of humans), they can and do easily visualize sheep as they count them.
perrygeo··on How bicycle coaster brakes work (2018)
> Where is all this love for coaster brakes coming from?!

Childhood nostalgia? They are indeed terrible brakes from a riding perspective. But a) no cables or pads to mess up, b) beginner friendly and c) skidding. Great for kids riding around the neighborhood, the only real purpose.

perrygeo··on Your AGENTS.md file doesn't do anything
Yes. Extrapolated further, my AGENTS.md files barely have any content. It's an index with links to code sections, skills, documentation, etc. I think of AGENTS file like a router - the job is present what's available and dispatch accordingly.
perrygeo··on Ask HN: Which class of business model is most resilient to build in the AI era?
> Most new businesses have some form of AI in them

Most new business are customers of AI companies. They are consuming AI. They don't build AI. They don't advance AI technology. They are not using AI in innovative ways. They are consumers.

Don't let anyone fool you into thinking there are a bunch of new AI companies around. No. There are a few AI companies, and millions of AI customers.

Hey look, I use electricity produced by someone else. I must be in the Energy sector!?

Hey look, I stock the fridge with veggies someone else grew. I must be innovating in Agriculture!?

Hey look, my product was created by a someone else's robot. I must be innovating in AI!?

Obviously not. Any valid conversation starts by differentiating between a leading innovator, a producer and a mere consumer.

perrygeo··on Aphantasia Beginner's Guide
Visualizing a sheep is not something I can even do once. Let alone repetitively until it gets boring. Being asked to visualize is quite frustrating and confusing - especially to a little kid not sure exactly what he's supposed to be "seeing". Hardly relaxing.
perrygeo··on Aphantasia Beginner's Guide
I thought "count sheep" just meant "please just go the f*k to sleep, kid". Literally counting sheep was too absurd. At the risk of sounding obvious, there are no sheep. No amount of trying to conjure pictures of sheep would work. I was very confused by this as a kid.

Later in life, the realization moment came when I was watching a TV show, a police procedural where the witnesses described the suspect. Police were able to use the descriptions to make an accurate sketch. "Yeah right, like anyone could recall those details". I was convinced it was a Hollywood trick.

My friends looked at me funny and proceeded to describe the faces, clothing, etc. of off-screen characters with stunning accuracy. That's when I learned other people can recall (or even conjure) visual images and read them like photographs.

perrygeo··on Fences, Not Sandboxes
I think we're in total agreement here. The $120k gets a profit for Anthropic. The $5k subscription loses money hand over fist. Somewhere in between is the real market price. And the OP is betting on a 10x reduction, from his already bargain-basement subsidized fee? Right at the moment when the market is demanding returns? Bold strategy Cotton, let's see how it plays out.
perrygeo··on Fences, Not Sandboxes
That $5k subscription is already deeply discounted by debt funding. Anthropic is losing money on that account. So the bet is this free lunch will continue indefinitely and prices are gonna drop 10x on top of that? We'll see!

Put another way, someone else is paying that 115k difference, and it's not out of the goodness of their hearts. Once they capture market share, capitalists gonna capitalize. For now, investors seem happy to pump money into the bubble without seeing returns, but that will not last forever.

perrygeo··on Oceans hit highest temperature on record
Yes, scientists have put forth dire predictions of extreme weather. Here's the thhing though - we actually see measurable increases in the frequency and severity of extreme weather events. They put forth dire predictions.. and they are absolutely correct. It's right there in the observational record.
perrygeo··on Fences, Not Sandboxes
> So it's not $120k/month of real money, but it's still crazy spend. I would guess I'm one of a handful of top individuals on Earth outside the frontier labs, in terms of my experience with top-end models.

We get it bro, you buy lots of tokens. My son buys lots of pokemon cards. Still haven't seen shit for return on that investment either, but as long as its all in good fun, "you do you".

> That's how I'm able to tell the future. I'm living in a world that will not become cost-effective for most people for another year.

That's pretty easy to verify. #remindme August 24, 2027 - are $120k worth of tokens in 2026 generally "affordable" by 2027? By what mechanism have the economics shifted?

perrygeo··on Anthropic's best AI model struggles to attract users as cheaper tools thrive
For research tasks, I like Meta's muse glimmer 30b model run locally. Whatever they did in post-training to make it so terse and direct, it worked. Gets right to the point and wastes no tokens. Here's a snippet from a recent thinking trace:

> Calculation done. Potential confusion: sq km vs acres. Provide both. Output.

Other models would have written paragraphs dancing around the idea. Glimmer barely does sentences.

It doesn't really sound human - which is a plus in my book. I want my robots to talk to me like they are robots.

perrygeo··on Stop Anthropomorphizing Intermediate Tokens as Reasoning/Thinking Traces (2025)
> next token prediction is not searching a solution space.

Interesting take. Next token prediction (via the attention mechanism) is a "walk" through the token embedding space. Searching the solution space is what it does, mathematically. It's how we take tokens x context length possible combinations and prune them to converge on viable answers so quickly.

Does it look like search at inference time? No. With given weights, a given prompt, and a given random seed, you get the exact same answer. There's not much searching happening at inference...

The key is that most of that space is searched at training time. The weights implicitly prune the search space, blocking off or make certain token combinations effectively impossible. It's easy to think "we're just applying weights at inference time" without considering all the pre-work that's done to prune that search space.

Which is exactly why "thinking" traces (and randomization) are useful! They bust out of any local optima created by too-tightly-constrained models or system prompts. It's both useful and technically correct to speak of the process as a high-dimension combinatorial search.

perrygeo··on A Friendly Introduction to Racket
Yep, `/usr/bin/gdb /usr/local/bin/clisp 1337` ... also in TADC lore, windows has forward slashes.
perrygeo··on A Friendly Introduction to Racket
> In The Amazing Digital Circus (episode 8, "hjsakldfhl"), when Kinger opens the terminal to try to reset Caine, you can see that Caine (a creative AI built in 1996) is programmed in Lisp. The file is literally named Caine-core.lisp.

Nice touch. Explains how Caine returns in episode 9: Lisp continuations allow for graceful error recovery.

perrygeo··on I should have loved biology (2020)
Skin in the game. Being "smart" isn't an immutable property of a noun. It's a verb. It's something you do, something you practice, constantly.

Even in fields like Math where ostensibly we have universal truths accessible to all, we still must learn how to derive proofs. Without doing math, one cannot understand math.

Learning IS experiential. It requires concentration and effort. Those who seek to make learning "easier" by removing effort are embarrassingly misguided about human cognition.

perrygeo··on Stop Anthropomorphizing Intermediate Tokens as Reasoning/Thinking Traces (2025)
Coming from a more traditional stats/ML background, I try to view "thinking" traces as a way to explore the search space without getting caught in a local maximum.

A better analogy for me is annealing; you can't cool metal down instantly or the result is brittle. You must cool down gradually, which allows the molecules to arrange into more durable structures. Random but controlled.

In the same way, thinking traces are testing out all sorts of novel connections between tokens ("But wait..", "Actually,.."). Like a highly divergent branching mind map that gets pruned over time, rather than settling directly into the initial answer.

Now consider human cognition. We're constantly diverging, daydreaming, playing "what if" scenarios and measuring up those ideas against our internal objective functions (proxy for reality) to see which ideas stick. Not too dissimilar. But it's hard to call what we do "thinking" either - it's the default mode network wandering.

perrygeo··on Don't Paste the AI, please
Pasting AI slop is a strong signal. It tells the world "I add no value. I am a proxy for ideas generated by others". It's sad to see people I otherwise respected debase themselves.

I'll say this clearly. Using AI for writing is bad. Don't do it. Ever. If you can't formulate a coherent thought yourself, just don't. I don't want to hear what the LLM said. I want to hear what you said. Loudly proclaiming you have nothing to add is a strange flex. Just stay quiet.

perrygeo··on 2,085 Tests, and None of Them Opens the Front Door
> Unit and integration tests serve different purposes and one isn't necessarily better than the other.

If they serve different purposes, then one is by definition better than the other, given the circumstances.

There's a whole class of applications containing very little pure logic; dumb pipes hooking together a database and client over a network. It's a categorical mistake to fully unit test these. There are only a few "units" to test; everything else is a side-effect (disk, network, etc). Yeah you can mock. My point is that mocks do not cover real world behavior, so those types of apps require an integration/e2e heavy set of tests.

Conversely, if you've got a logic-heavy library, you need unit tests. You can't rely on the behavior as filtered through the application.

Different purposes. But depending on what world you're working in, one is definitely better than the other.

perrygeo··on Abdominal fat predicts heart disease risk better than BMI
Visceral fat really is the devil. It's not some inert glob of tissue padding your organs, it's metabolically active in all the wrong ways - increasing inflammation and insulin resistance.

I've stopped worrying about weight (or BMI) and focus solely on waist circumference as a proxy for visceral fat.

perrygeo··on The tokenmaxxing hype didn't last long
You can't assume more tokens == more value. Turns out bad ideas, amplified, just turn in to bad implementations.

Github says we've 14x the number of commits this year. There is no evidence of anything close to 14x economic value being created. Most of those tokens are purely waste.

← PreviousPage 3 of 34Next →