HNHacker News
TopNewBestAskShowJobs

tikhonj

16,934 karma · joined September 12, 2011

Tikhon Jelvis

I am interested in programming languages, functional programming (especially Haskell), domain-specific languages, static analysis and program synthesis.

I am working on apply programming language ideas to supply chain planning.

I worked on supply chain optimization at Target for several years, and co-authored a textbook on reinforcement learning based on that experience.

contact: http://jelv.is, tikhon@jelv.is

submissionscomments
tikhonj··on Anthropic's Method to Losing Goodwill in a Few Easy Steps
I mean, it has definite Borg Bill Gates vibes, doesn't it?

It's totally reasonable for tech folks to be leery of a company using a strong market position from one system (Windows/Claude) to push mediocre complementary software (IE/CC).

tikhonj··on It's not about physical vs. digital games, it's about ownership
Which online stores? I don't see that in Steam.
tikhonj··on California bans 'sell by' labels, hoping to cut food waste
Regulations aren't fungible, so this is not a particularly meaningful perspective.
tikhonj··on Synthesis is harder than analysis
It's an illustration for a blog post on a largely different subject, it really doesn't matter.

Who would be looking up the integral of one of the most common functions in applied math in a random philosophical article aimed mostly at SREs?

tikhonj··on HackerRank open sourced its ATS. My resume scored 90/100. Oh wait 74. No – 88
The US has a notion of "disparate impact"[1] that means you can be liable for discriminating based on a protected characteristic on the basis of correlation. This is why HR departments are very hesitant to use things like IQ tests for screening candidates, for example.

[1]: https://www.congress.gov/crs-product/IF13057

tikhonj··on You Don't Know Jack About Formal Verification
Property based testing is useful for finding bugs even in these kinds of CRUD heavy apps. There can be a surprising number of bugs and unexpected behaviors in the interaction of multiple sub-systems or APIs, and one way to find those bugs is to come up with properties on traces of calls.

For example, I saw a paper on using metamorphic testing (a particular technique for defining properties to test) to find bugs in the web APIs of Spotify and YouTube[1]. I don't have time to reread the paper just now, but I remember that they found weird behavior in pagination of search results. Not a big deal in that particular case, but I've definitely seen internal APIs with behavior that could be similarly wrong but with a much larger real-world impact.

[1]: https://ieeexplore.ieee.org/document/8074764

Personally, I see property-based testing and formal specification more as tools for design and debugging more than full-on correctness. Even with AI models it's still really hard to fully prove something correct, but having even a partial logical specification can let you trade design time for debugging time and lets you find inconsistencies or potential edge-cases when you're initially figuring out a system, rather than waiting until you have a massive codebase to debug and redesign in production.

It's not a panacea and you still have to be careful at the boundary between your nicely modeled system and the real world, but, once I got the hang of working in that style, having some formal properties or partial logical specifications of the behavior I needed has been really nice to have, as much for saving effort as for ensuring correctness.

I've mostly worked in slightly different domains, but I've found property-based testing useful both as a tool to catch bugs but also as a tool for communication. I spent a couple of years building and supporting a supply chain simulation at Target, where I frequently got requests for new metrics from the supply chain planning team. By teaching them how to specify either the whole metric or, at least, some of the expected behaviors of the metric as mathematical properties, it took far fewer back-and-forth conversations to correctly implement the metric in the simulation. We could then test these things by checking these properties over a bunch of random simulation traces. Day-to-day this saved a bunch of time in debugging small simulation mistakes. In the longer-term, having this test suite also let us rewrite the simulation code in a new style in Rust to drastically increase performance. All of this would have been possible without the set of properties, it would have just involved a whole lot more slow, tedious work.

tikhonj··on The Garbage Collection Handbook: The Art of Automatic Memory Management (2nd Ed) (2023)
Ironic that this of all books got garbage collected prematurely.
tikhonj··on Emacs, how it all started for me
The neat thing with Emacs is that the core concepts of the system are all first-class programming entities with their own documentation. So if you want to know what your current mode does, you can use C-h m to get a bunch of information including commands, key-bindings and links to code. If you have a key command, you can use C-h k, enter the keys, and you'll see exactly what function that command runs. You can get info about functions with C-h f and variables with C-h v; coupled with some kind of fuzzy-find-autocomplete (which, unfortunately, isn't set up by default), it's usually pretty quick to find the functions and config options that are relevant to whatever you're trying to do.

I still use web searches to look up Emacs things occasionally, but the built-in help commands are still useful because they're naturally tied to (and organized by) the core code entities that power Emacs.

tikhonj··on FTX's former Anthropic stake would be worth about $75B at today's valuation
So, if FTX had managed to stay afloat a bit longer, they could have gotten away with it.

Which is a much more interesting statement in the context of certain other crypto organizations than it is about FTX...

tikhonj··on How to earn a billion dollars
"Let's have more world wars" isn't a great thing to aim for.
tikhonj··on Nobody ever gets credit for fixing problems that never happened (2001) [pdf]
Counterpoint: I absolutely give credit to Sonic for being a great ISP and recommend them to everyone. I got my parents to switch when Sonic finally rolled out to their neighborhood.

If online comments are anything to go by, I'm not alone.

If you're in the Bay Area and you can get a Sonic fiber connection, I would highly recommend them over AT&T/Comcast/etc.

tikhonj··on Emacs appearances in pop culture
You always have a choice. Sometimes the best move is not to play.
tikhonj··on Shall we play a game? My AI nuclear simulation
The point is that it's the same process with—much—better priors.

This seems like a reasonable view to me. It's surprising just how much better priors matter and how we can develop those priors by training on a bunch of text. But it also explains, or at least hints at an explanation, for why LLM capabilities are so jagged, and in such inhuman ways.

tikhonj··on H2JVM – A Haskell Library for Writing JVM Bytecode
This project isn't for running Haskell on the JVM, it's for writing a compiler that produces JVM bytecode. You'd use it if you wanted to implement your own JVM language in Haskell or maybe if you wanted to have some kind of JVM-backed domain-specific language embedded in Haskell.
tikhonj··on My Software North Star
Just because you can't get everyone to agree does not mean that the concept is not well-defined or clear. People can and do disagree over almost anything. That could just mean that one side is wrong and the other side is right, even if it is difficult to determine which is which... or simply difficult to navigate the underlying politics regardless.

Besides, in your example, either kind of gap could be a bug or a missing feature. It's a totally orthogonal question.

tikhonj··on My Software North Star
There is a pretty clear and important difference between a program that does something wrong, and a program that just doesn't do something somebody wants.

"You don't support printing", "pressing the print button doesn't print", "pressing the print button crashes the computer" and "pressing the print button lets an attacker get root access to the system" are all different and it makes sense to distinguish them. (The first is a missing feature, the second and third are different kinds of bugs, the last is a special kind of bug we call a security vulnerability.)

That distinction might not be useful to end-users, but it's useful for the people building the system! If you want to care about quality, committing to a strategy like "we will not add features before we fix known bugs" is totally clear, reasonable and effective. There might be some frontier of issues where it's hard to make a distinction, but that just means there are subtle edge-cases, not that the whole concept is undefined. A lot of perfectly cromulent concepts have edge-cases! You can just decide those on a case-by-case basis; if it's actually so close as to be legitimately confusing—it's not just feigned ignorance or political posturing—which side you choose probably doesn't have much of an effect.

This does depend on having a reasonably clear idea of what you're building, but that "reasonably clear idea" does not have to be anywhere near the detail of a "full spec", much less anything formalized. To me, that seems like a baseline you'd need to build quality software at all, and hardly an unreasonable thing to expect. And if most teams can't manage, well, it's just another explanation for why most software is crap.

tikhonj··on DNS is for people, not for IT infrastructure
DNS is a database.
tikhonj··on Uber's $1,500/month AI limit is a useful signal for AI tool pricing
I still believe Scrum is a fad and yet companies have been spending obscene amounts on to push it down developers' throats for decades now.
tikhonj··on Can the stockmarket swallow Anthropic, SpaceX and OpenAI?
Because they think that a lot of people will want to get in on the historically massive and well-known companies, which would lead to outflows if the index doesn't pick them up fast enough?
tikhonj··on macOS needs its grid back
That seems ≈impossible in a world where you're running arbitrary, Turing-complete code. A modern consumer machine can do so many different things—often a bunch at a time—that there is always a massive amount of space to hide bad behavior.

There might be some way to design a system from the ground up to avoid this problem (some kind of declarative, capability-based security?), but retrofitting that onto an existing behemoth of a system does not really work.

tikhonj··on Shift from a leader-follower to a leader-leader approach
Leaders can matter—a lot—even if they do not exercise granular, top-down control. They can provide the right context and motivation, articulate high-level aims, resolve issues and create the kinds of systems and cultures that do not need explicit direction.

A great illustration of this idea in context is Andrew Gordon's book on the Battle of Jutland, The Rules of the Game[1]. He mostly contrasts the leadership philosophies in the British Navy shortly before and during the Battle of Jutland. The British Navy became very top-down and procedure-oriented during peacetime, which did not generalize well to operating in battle.

[1]: https://www.goodreads.com/book/show/354137.Rules_of_the_Game

tikhonj··on What if remote working, not AI, is to blame for weak junior hiring?
What if it's interest rates?

I'm amazing at how anybody can discuss any questions about the tech job market without addressing the end of zero-interest rate policy. "It's interest rates, stupid" ought to be the first thing we consider, and stuff about AI/remote/etc only comes in when we have a specific reason to consider it.

tikhonj··on Shift from a leader-follower to a leader-leader approach
Part of the idea—which requires non-trivial effort from the leaders to build up sufficient trust—is that people should believe that their proposals (for their own work) are accepted by default. People don't have to wait for their idea to be accepted, but they have to communicate enough so that leaders can surface specific issues or mismatches only if they have them.
tikhonj··on Shift from a leader-follower to a leader-leader approach
Wartime is exactly when centralized control breaks down the hardest, because conditions start changing incredibly fast and communication breaks down. There's a reason the phrase is "fog of war" and not "fog of peace"!
tikhonj··on Pandoc Templates
WYSIWYG pretty consistently leads to visual and structural messes. It's only going to "beat" everything else if you don't care about quality.

Most people don't—and don't have to—care about quality for their short, simple documents, but that is neither good nor inevitable, and it's always worth trying to do better.

tikhonj··on Shift will clean homes for free to train future robots
Haha, yes, meant to write "hotel management". I'll update the answer to fix that.
tikhonj··on I am retiring from tech to live offline
Living in a cookie-cutter suburb full of parking lots and strip malls is not, in fact, a higher standard of life than living in a small town in Europe. The fact that we've somehow convinced ourselves that it is says absolutely nothing good about American postwar culture.
tikhonj··on Shift will clean homes for free to train future robots
My guess is that they're currently nowhere near robust or effective enough to make that realistic. They need to bootstrap somehow, if only get good enough to convince hotel management that their approach will be realistic in the future.
tikhonj··on I am retiring from tech to live offline
I broadly agree, but expensive homes are an odd example because, like, you can just sell the house later. There might be some opportunity cost relative to investing in a broad market ETF or whatever, but that's not the sort of difference that makes or breaks your early retirement strategy.

It's a bit odd to get moralistic over saving/spending money in general, but that's especially true around expensive homes.

tikhonj··on Garnix (A Nix CI) is shutting down
Presumably it was and they got acqui-hired.
← PreviousPage 4 of 34Next →