HNHacker News
TopNewBestAskShowJobs

criemen

395 karma · joined June 28, 2025

submissionscomments
criemen··on Supply chain attacks are exploiting our assumptions
Fair enough, it makes more sense in a, say, Haskell-style pure functional language. Instead of getting a general IO monad, you pass in more restricted functionality.

Still, it'd be highly painful. Would it be worth the trade-off to prevent supply chain attacks?

criemen··on Supply chain attacks are exploiting our assumptions
> we should instead isolate every library we import and run it under a sandbox

I don't see how that'd be possible. Often we want the library to do useful things for the application, in the context of the application. What would incentivize developers to specify more fine-grained permissions per library than the union of everything their application requires?

I see more use in sandboxing entire applications, and giving them more selective access than "the entire user account" like we do these days. This is maybe more how smartphones operating systems work than desktop computers?

criemen··on Tongyi DeepResearch – open-source 30B MoE Model that rivals OpenAI DeepResearch
It's hard for me to say. I don't think you know you're on the S-curve until after the fact.

On the one hand, most models are "good enough" for chatgpt-like usage, and there it's hard to see/feel generation-to-generation improvements. On the other hand, if you look at instruction following, dealing with long context windows, >200 tool call interactions while staying on track, there's still plenty of improvements to be had. So, hard to say where we are.

criemen··on Tongyi DeepResearch – open-source 30B MoE Model that rivals OpenAI DeepResearch
I tend to use them when I'm looking to buy something of category X, and want to get a market overview. I can then still dig in and decide whether I consider the sources used trustworthy or not, and before committing money, I'll read some reviews myself, too. Still, it's a speedup for me.
criemen··on Tongyi DeepResearch – open-source 30B MoE Model that rivals OpenAI DeepResearch
> or if it takes a couple of months to fold these advantages back into the frontier models.

Right now, I believe we're seeing that the big general-purpose models outperform approximately everything else. Special-purpose models (essentially: fine tunes) of smaller models make sense when you want to solve a specific task at lower cost/lower latency, and you transfer some/most of the abilities in that domain from a bigger model to a smaller one. Usually, people don't do that, because it's a quite costly process, and the frontier models develop so rapidly, that you're perpetually behind them (so in fact, you're not providing the best possible abilities).

If/when frontier model development speed slows down, training smaller models will make more sense.

criemen··on Why does Swiss cheese have holes?
In German, "Swiss cheese" is a term that's well known, and doesn't count that kind of amusement.

For example, you could say that something "looks like swiss cheese" when it has a lot of holes in it, like very old clothing. It's often used slightly ironic, but that's not due to what you state.

criemen··on At the end you use `git bisect`
+1: EdaMagit has been a game changer for me wrt reordering commits, fusing them together, and writing at least 1-2 sentences of proper commit messages after the fact.
criemen··on Composer: Building a fast frontier model with RL
I understand where you're coming from, and I'd love to have learned about pre-training vs. off-the-shelf base model too. But

> their own internal benchmark that they won't release

If they'd release their internal benchmark suite, it'd make it into the training set of about every LLM, which from a strictly scientific standpoint, invalidates all conclusions drawn from that benchmark from then on. On the other hand, not releasing the benchmark means they could've hand-picked the datapoints to favor them. It's a problem that can't be resolved unfortunately.

criemen··on The last European train that travels by sea
This is not true, on the route Denmark/Germany there was no decoupling involved - the train drives onto the ferry, and drives off the ferry afterwards.
criemen··on The last European train that travels by sea
No - it's a rather short train, only a few carriages, and the engine is mostly integrated into the carriages I believe.
criemen··on The last European train that travels by sea
> I doubt that they carry the engine across by ferry?

I don't know about this specific instance, but the Germany-Denmark train-on-a-ferry carried the whole Diesel-powered train, engine and all. It drove onto the ferry on its own, and left it on its own as well.

criemen··on Mistakes I see engineers making in their code reviews
Great article, and I mostly agree.

I want to nitpick on

> Don’t review with a “how would I write it?” filter

_Thinking_ about "how would I write it?" when reviewing code catches a lot of stuff for me - I want to be able to come up with how I'd do it, see how the PR does it, and then check if the approach matches. If not, there's something to be learned and explained here. Although, all of that is in my head! If I come to the conclusion that the presented approach is wrong or inferior in an objective way (say, doesn't handle corner cases neatly or at all), I'll comment. Otherwise, let people do their thing!

So, I agree with the point of not using this to generate a lot of comments, but I do believe in it as a useful technique to engage the brain, instead of just loosely scrolling over the diff.

criemen··on Microsoft Teams will start tracking office attendance
I don't think macbooks have a GPS chip built-in, isn't it only wifi geolocation?
criemen··on Meta is axing 600 roles across its AI division
How should they do that? I hear that phrase, and it's easy to agree to, but how would it look in practice?
criemen··on The case for the return of fine-tuning
> They will hire anyone who can produce a model better than GPT5, which is the bar for fine tuning

Depends on what you want to achieve, of course, but I see fine-tuning at the current point in time primarily as a cost-saving measure: Transfer GPT5-levels of skill onto a smaller model, where inference is then faster/cheaper to run. This of course slows down your innovation cycle, which is why generally this is imo not advisable.

criemen··on AI has a cargo cult problem
> Supposedly the performance of Owen-coder is comparable to the likes of Sonnet4. If I invest in a homelab that can host something like Qwen3 I'll recoup my costs in about 20 months without having to rely on Anthropic.

You can always try it via openrouter without investing in the home setup first. That allows you to evaluate whether it hits your quality bar or not, and is much cheaper. It is less fun than self-hosting though.

criemen··on Claude Haiku 4.5
But that doesn't make sense? Why would they keep the cache persistent in the VRAM of the GPU nodes, which are needed for model weights? Shouldn't they be able to swap in/out the kvcache of your prompt when you actually use it?
criemen··on Claude Haiku 4.5
Fascinating, so I have to think more "pay for RAM/redis" than "pay for SSD"?
criemen··on Claude Haiku 4.5
Technically, they released Opus 4.1 a few weeks ago, so that alone hints at a smaller leap from 4.1 -> 4.5, compared to the leap from Sonnet 4 -> 4.5. That is, of course, if those version numbers represent anything but marketing, which I don't know.
criemen··on Claude Haiku 4.5
> I suppose it depends on how you are using it, but for coding isn't output cost more relevant than input - requirements in, code out ?

Depends on what you're doing, but for modifying an existing project (rather than greenfield), input tokens >> output tokens in my experience.

criemen··on Claude Haiku 4.5
I don't understand why we're paying for caching at all (except: model providers can charge for it). It's almost extortion - the provider stores some data for 5min on some disk, and gets to sell their highly limited GPU resources to someone else instead (because you are using the kv cache instead of GPU capacity for a good chunk of your input tokens). They charge you 10% of their GPU-level prices for effectively _not_ using their GPU at all for the tokens that hit the cache.

If I'm missing something about how inference works that explains why there is still a cost for cached tokens, please let me know!

criemen··on C++26: range support for std:optional
Great to see they learned from Java, which initially made the mistake of not supporting the streams interface for their Optional type at first [1]. It was infuriating.

[1] https://stackoverflow.com/questions/22725537/using-java-8s-o...

criemen··on Finding a VS Code Memory Leak
> For obscure technical reasons they are always multiples of four.

Why did I know that that link went to Raymond Chen before clicking it... That man is a treasure.

criemen··on LLMs are mortally terrified of exceptions
It's also logically incoherent - division by zero can't occur, because if b=0 then abs(b) < sys.float_info.epsilon.

Furthermore, the code is happy to return NaN from the pre-checks, but replaces a NaN result from the division by None. That doesn't make any sense from an API design standpoint.

criemen··on A small number of samples can poison LLMs of any size
> It is well-known that even the biggest SOTA models require only 100-200 good samples for fine-tuning.

As someone who's not heard of this before, do you have a link for this? Is this LORA-finetuning only? Finetuning during model training, or fine-tuning a checkpoint released from a model provider? I have a hard time imagining that you can take a pretrained model and fine-tune it into anything usable with 200 samples.

criemen··on A small number of samples can poison LLMs of any size
> Please provide a citation for wild claims like this. Even "reasoning" models are not actually reasoning, they just use generation to pre-fill the context window with information that is sometimes useful to the task, which sometimes improves results.

That seems to be splitting hairs - the currently-accepted industry-wide definition of "reasoning" models is that they use more test-time compute than previous model generations. Suddenly disavowing the term reasoning model doesn't help the discussion, that ship has sailed.

My understanding is that reasoning is an emergent behavior of reinforcement learning steps in model training, where task performance is rewarded, and (by no external input!) the model output starts to include phrases ala "Wait, let me think". Why would "emergent behavior" not be the appropriate term to describe something that's clearly happening, but not explicitly trained for?

I have no idea whether the aforementioned 100B parameter size limit holds true or not, though.

criemen··on One-man campaign ravages EU 'Chat Control' bill
Yes, yet [1] happened where someone took precautions to drive safely at >400 km/h speed.

[1] https://www.drivencarguide.co.nz/news/no-charges-laid-on-bug...

criemen··on Vibe engineering
I wanted to point you at https://neverworkintheory.org/ which attempted to bridge the gap between academia and software engineering. Turns out the site shut down, because (quoting their retrospective)

> Twelve years after It Will Never Work in Theory launched, the real challenge in software engineering research is not what to do about ChatGPT or whatever else Silicon Valley is gushing about at the moment. Rather, it is how to get researchers to focus on problems that practitioners care about and practitioners to pay attention to what researchers discover. This was true when we started, it was true 10 years ago, and it remains true today.

The entire retrospective [1] is well worth a read, and unfortunately reinforcing your exact point about software development being fashion/emotion driven.

[1] https://www.computer.org/csdl/magazine/so/2024/03/10424425/1...

criemen··on Indefinite Backpack Travel
If only the macbook opened more than the 120 degrees (or so) angle with the screen I'd be sooo happy already (no touch required).
criemen··on The Theatre of Pull Requests and Code Review
I agree with you that this shouldn't be 100% hard defaults, but it's a good standard to have, and imo it's valuable to be explain why one is deviating from it.

> - smaller PRs aren’t necessarily easier to review (and this arbitrary obsession almost always leads to PR overload in chunks that don’t make any sense, reducing code quality as a result)

Oh but they sure can be reviewed more easily, because they are shorter? Doing so feels like less effort, and you get a dopamine hit from hitting that "submit review" button faster/more often (improved morale, and PR turnaround time!). Plus, if there's a longer discussion about X, it's great if it's not tangled up with Y and Z at the same time - allowing you to dig into X.

> - nobody reads intermediate commit messages one by one on a PR, period.

Come on, that's intellectually dishonest. 1. VSCode displays commit messages inline as blame for me (and many of my colleagues), so even when we don't read the commit messages one by one _on a PR_, I often read them later in the IDE (we don't squash merge PRs). I spend significantly more time reading code than writing, and commit messages, PR descriptions and linked issues provide extra context that is useful to me especially for complex code. If those messages were entirely unreadable, I'd be annoyed. 2. When someone invests time into telling a good story commit by commit, in my team they write "Review commit-by-commit is encouraged" in the PR description, to tell the reviewers that yes, they should read the individual commits, as that'll make understanding the PR easier. Often as reviewer, I follow that suggestion.

> Wasting your time with this in a branch, as you work towards a solution, is focusing on the wrong thing

It seams you're conflating "working on a feature" with "presenting it as PR to review". That's two very different things, and Edamagit in VSCode makes it so so easy to provide a reasonable commit history that hides some of your missteps, and to fill in commit messages.

← PreviousPage 4 of 5Next →