vjerancrnjak

Great find and write up. This year if I remember correctly 40+ people got the check, ~1000 have an account at the bank. I got mine this year, an exercise I revisit every few years since 2012 to learn a new programming lang or approach had a newer update that made it have 2 offbyone errors. I had extra time this year so went to the beginning of the chapter to attempt an open problem and in the preliminaries another off by two error. I was quite surprised.

Really made me appreciate how unlikely it is to find an error. It feels as if it was planned just for me to find it. Just like the author studied cryptography and then decided to do some exercises to hone his skills, an unlikely journey towards a check.

Nope. Codex formalizes much better than any tool with exception of Aristotle from Harmonic.

https://github.com/vjeranc/fixed-rtrt

M3 module was formalized fully purely from experimental data and from a nudge by earlier versions of codex in 15-30 minutes in a simple write/compile/fix-first-error loop. I was a bit surprised how fast it picked up the pattern but given there was a paper from '70s it became clear why later.

David Attenborough saw more clearly than most what was being lost. But even he stopped short of fully applying that logic to animals themselves.

Rewilding at scale, deep emissions cuts, and a serious move away from animal agriculture are the same project.

Regret analysis in bandit and similar algorithms shows how inference is connected to loss function. If your loss function is good, greedy inference is as good as joint inference.

Training on cost-to-go loss is good enough. Perfect cost-to-go eliminates the need for global algorithms and allows local decision making. Given “natural” datasets it is probably the best thing to attempt to learn. The fact that probabilistic graphical models never really worked proves it somewhat.

Github was already struggling with bazillions of throw-as-much-crap-on-the-wall software running in actions, and now the world is running throw-as-much-LLM-crap-on-the-wall computation, as unstoppable as the pre-LLM era. Turning compute into excrement as fast as the planet is filled with it. Excrement being "Github Copilot code review" in compute world, and no need to draw what it is in our real world.

Weird that Anthropic decided to build a Claude Code Routines toilet.

Ask it to formalize it in Lean.

I think this description is often associated with ADHD memes.

Falling asleep after a can of energy drink.

Following the bottle-cap madness, I don't think any current data shows the actual issue was resolved. Even worse, the effect on marine life is still not measured, and afaik reduction of harm was the primary goal. Instead of brutally high fines on fishing net waste, we got bottle-cap madness.

We have so much experience with scientific method, yet these massive decisions are adhoc, that's how the whole world works. We never tested what would happen by allowing mass production of plastic, or phones, or whatever, so these antipatches are going by the "feels" as well, with no individual taking responsibility for failures.

just keep-alive it with pipelining, depending on the server, 100k+ RPS.

haha, all of a sudden I see a tab "waifu pillow" on Amazon, and think I have a split personality that runs searches in between consciousness shifts, and then I come back to a funny message.

Just asked for boxes, curved shapes, cylinders, cuts, mirroring, gave exact dimensions and it built up to ~400 lines of FreeCAD python.

3d printed handle was exactly how i wanted it to be

I vibecoded a suitcase handle months ago with its Python interface. A pleasant experience.

Your “what about plants” argument is such a worn-out trope that you must have seen it before and read a valid explanation of why it makes no sense.

Peter Singer has been writing on the topic for decades, including others. What-about-plants needs to fade away.

I had a similar feeling trying to calculate some combinatorial structures. At some point the LLM made a connection to extremal combinatorics and calculated tighter bounds and got me to the solution faster.

Felt flashbacks of playing chess against humans online as a teen by copying moves from a chess engine.

Whats the point haha

Libraries create boundaries, which are in most cases arbitrary, that then limit the way you can interact with code, creating more boilerplate to get what you want from a library.

Abstractions are the source of bloat. Without abstractions you can always reduce bloat, or you can reduce bloat in your glue, but you can't reduce glue.

It takes discipline to NOT create arbitrary function signatures and short-lived intermediate data structures or type definitions. This is the beginning of boilerplate.

So many advances in removing boilerplate are realizing your 5 function calls and 10 intermediate data structures or type definitions, essentially compute a thing that you can do with 0 function calls and 0 custom datatypes and less lines of code.

The abstraction hides how simple the thing you want is.

Problem is that all open source code looks like the bloat described above, so LLMs have no idea how to actually write code that is without boilerplate. The only place where I've seen it work is in shaders, which are usually written to avoid common pitfalls of abstraction.

LLMs are incapable of writing a big program in 1 function and 1 file, that does what you want. Splitting the program into functions or even multiple files, is a step you do after a lot of time, yet all open source looks nothing like that.

White colar work is just a lucky place to be, 99% of it is completely made up, there's people doing nothing, and people doing work of 10 people, does not matter, the work itself has no impact on anything.

A nice way to realize why this AI wave hasn't produced massive economy growth, it is mostly touching parts of economy which are parasitic and can't really create growth.

No. There is good signal in IMO gold medal performance.

These models actually learn distributed representations of nontrivial search algorithms.

A whole field of theorem provingaftwr decades of refinements couldn’t even win a medal yet 8B param models are doing it very well.

Attention mechanism, a bruteforce quadratic approach, combined with gradient descent is actually discovering very efficient distributed representations of algorithms. I don’t think they can even be extracted and made into an imperative program.

On reddit it's even worse, I feel like Reddit is internally having their own bots for engagement bait.

As someone who loves LaTeX, I can't imagine ever spending so much time on typography on online forums, italics, bold, emdashes, headers, sections. I quit reddit and will quit hn as well if situation worsens.

Funny how pydantic is used to parse and not validate but then there are post conditions after parsing which you should parse actually or which can be enforced with json schema and properly implemented constrained sampling on the LLM side.

Pipelines are often dynamic, how is this achieved?

Pipelines are just a description of computation, sometimes it makes sense to increase throughput, instead of low latency, by batching, is execution separate from the pipeline definition?

It's quite tricky as they optimize the agent loop, similar to codex.

It's probably not enough to have answer-prompt -> tool call -> result critic -> apply or refine, there might be a specific thing they're doing when they fine tune the loop to the model, or they might even train the model to improve the existing loop.

You would have to first look at their agent loop and then code it up from scratch.

I would say IMO results demonstrated that. Silver was tiny 3B model.

All of our theorem provers had no way to approach silver medal performance despite decades of algorithmic leaps.

Learning stage for transformers has a while ago demonstrated some insanely good distributed jumps into good areas of combinatorial structures. Inference is just much faster than inference of algorithms that aren’t heavily informed by data.

It’s just a fully different distributed algorithm where we can’t probably even extract one working piece without breaking the performance of the whole.

World/word model is just not the case there. Gradient descent obviously landed to a distributed representation of an algorithm that does search.

Most software is just middlemen collecting money. Thats the reason why there is no economic drastic growth even with 2-3 years of AI.

What you’re worrying about is just bigger middlemen.

If you were not a value extracting middleman there would be no fear of replacement, because you can always create more than what you take.

I’m glad if this causes a shift in the industry and we lose x analysts, x architects, x scientists, data engineers and all other formulaic middlemen that just live in a weird middlemen economy.

It is immense luck that we don’t have to actually produce something and we get paid but it is much better if we’re forced to actually do something that isn’t empty.

Bandits?

Spaced repetition algos

Flink is not really a performance choice, it's bloat to throw software as fast as possible at problems. I don't think there's any benchmark demonstrating insane capabilities per machine. I definitely couldn't get it to any numbers I liked, given other stream processing / state processing engines that exist (if compute and inmemory state management is the goal). Pretty sure any pathway that touches RocksDB slows everything down to 1-10k events per second, if not less.

The problem of finding out which video is next, by immediately taking into account the recent user context (and other user context) is completely unrelated to what Flink does -- exactly-once state consistency, distributed checkpoints, recovery, event-time semantics, large keyed state. I would even say you don't want a solution to any of the problems Flink solves, you want to avoid having these problems.

Flink is too slow for this.

If by features you mean tracking state per user, that stuff can be tracked without Flink insanely fast with Redis as well.

If you re saying they dont have to load data to update the state, I dont see how massive these states are to require inmemory updates, and if so, you could just do inmemory updates without Flink.

Similarly, any consumer will have to deal with batches of users and pipelining.

Flink is just a bottleneck.

If they actually use Flink for this, its not the moat.

I tried just repeating guó for as many times as symbols and repetition was not recognized.

Although I like the active aspect of the approach. Language apps where sound is the main form of learning should have a great advantage, as any written text just confuses as every country has its own spin on orthography. Even pinyin, despite making sense, for a beginner, has so many conflicting symbols.

I think this is standard. It applies to domains as well. I experienced government services blocks as well -- they send me an email, yet block my reply. I complain every time and rarely does anyone care, the support person does not escalate, so my email remains blocked, sometimes I'm told system is working as configured, completely ignoring that I am a real person and system is hostile towards me.

It's just general fragility of tech and lack of care from the creators/maintainers. These systems are steampunk, fragile contraptions that no one cares to actually make human friendly or are built on crappy foundations.

I did not do anything wrong. I had no choice with Redshift and had instructions from above. I made it work really well for what it can do and was surprised how much it sucks even when it has its own data inside of it and has to do compute. As a completely closed system, it's not impressive at all. It has absolutely shameful group-by SQL, completely inefficient sort-key and compression semantics, and absolutely can't attach itself to Kinesis directly without costing you insane amounts of money, because as you already know, Redshift is not a live service (you won't use it by connecting directly to it and expect good performance), it's primarily a parallel compute engine.

Your assessment of me is flawed. You haven't really shown any kind of low-level expertise on how actually these systems work, you've just name dropped OLTP OLAP as if that means anything at all. What is Timescale (now TigerData), OLTPOLAPBLAPBLAP? If someone tells you to use Timescale, you have to figure out how to use it and make the system yield to your will. If system sucks, it yields harder, if system is well designed, it's absolutely beautiful. For example, I would never use Timescale as well, yet you can go on their page and see unicorns using it. I have no idea why, but let them have their fun. There's successful companies using Elasticsearch for IoT telemetry, so who am I to argue I wouldn't do that as well.

There's nothing wrong with using PostgreSQL for timeseries data, you just need to know how to use it. At some point, scaling wise, it will fail, but you're deciding on tradeoffs.

So yes, my assessments have a good track record, not only of myself, but of others as well. I am extremely open to any kind of precise criticism and have been wrong bazillion times and I take part in these kinds of passionate discussions on the internet because I am aware I can absolutely be convinced of the other side. Otherwise, I would have quit a long time ago.

Clickhouse is column based storage, I can also apply delta compression, where gapless timestamp columns basically have 0 storage cost. I can apply Gorilla as well and get nice compression from irregular columns. I am aware of Redshift's AZ64 cols and they are a let down.

I can change sort order, same as in Redshift with its sort keys, to improve compression and compute. Redshift does not really exploit this sort-key config as much as it could.

My own assessment is that I'm extremely skilled at making any kind of DB system yield to my will and get it to its limits.

I have never used Redshift, Clickhouse or Snowflake with 1 by 1 inserts. I have mentioned S3 consumers (a library or a service, optimized to work well with autoscaling done by S3, respecting SlowDown -- something Redshift itself is incapable of respecting -- and achieving enormous download rates -- some of the consumers I've used completely saturate the 200Gbps limits of some EC2 machines at AWS). These consumers cannot be used in a 1-by-1 setting, the whole point is to have an insanely fast pipelining system with batched processing, interleaving network downloads with CPU compute, so that in the end, any kind of data repackaging and compression is negligible compared to download, so you can just predict how long the system will take to ingest by knowing what your peak download speed is, because the actual compute is fully optimized and pipelined.

Now, it might just be Redshift has bugs and I should report them, but I did not have the experience of AWS reacting quickly to any of the reports I've made.

I disagree, it's not a me problem. I am a bit surprised after all I've written that you're still implying I want OLTP, am using the wrong tool for the job. There are just some tools I would never pick, because they just don't work as advertised, Redshift is one of them. There are much better in-memory compute engines that work directly with S3, and you can create any kind of trash low-value pipelines with them, if you reach mem limits of your compute system, there are much better compute engine + storage combos than Redshift. My belief is that Redshift is purely a nontechnical choice.

Now, to steelman you, if you're saying:

* data warehouse as managed service,

* cost efficiency via guardrails,

* scale by policy, not by expertise,

* optimize for nontechnical teams,

* hide the machinery,

* use AWS-native bloated, slow or expensive glue (Glue, Athena, Kinesis, DMS),

* predictable monthly bill,

* preventing S3 abuse,

* preventing runaway parallelism,

* avoiding noisy-neighbor incidents (either by protecting me or protecting AWS infra),

* intentionally constrained to satisfy all of the above,

then yes, I agree, I am definitely using the wrong tool but as I said, if the value proposition is nontechnical, I do not really care about that.