omcnoe

Would we say that about any other industry? Can you run a film studio if you have never watched a film in your life? Can you run a publishing house if you’ve never read a novel?

It's competing with Starlink in that market, which is a much stronger product today.

For a project like Go, wouldn't it be an option to do one-time iouring feature detection in the runtime startup? Exploits are an issue for the entire OS, not the program choosing to use iouring, yeah?

LLMs can be pretty conmpetent at languages that have zero training data, at least to the extent that those languages use features/ideas that are familiar. I wrote a toy language/compiler and AI can write code for it competently.

Small customers grow into bigger ones later on. At least they do in the US, maybe this doesn’t happen so much in Adyen’s part of the world?

Scope is crucial when working with multiple teams/projects in a monorepo.

/mnt/c is a mounted C: drive in WSL2, that allows WSL2 guests to read/write files on the Windows host.

The mount is fine and speedy enough, but the underlying reads/writes turn into native NTFS reads/writes through Windows. NTFS file API is incredibly slow - high fixed overhead for initial file access.

So patterns like node_modules with many small individual files (or compiling code in general) are much much slower on Windows or WSL2 /mnt/c due to the fixed overhead adding up over a large number of files.

It's a ridiculous problem that has plagued Windows for years.

Outsourcing also fails because it’s a pathological case of adverse selection. The businesses that outsource projects are ones who are organisationally incapable of managing those projects well internally. But, that inability extends to their oversight of outsourcing shops as well.

End result is that many outsourcing firms are borderline fraudulent in the way they treat their customers.

Intelligence agencies use parallel construction to disguise their real methods. Further, more sophisticated methods are reserved for bigger targets. Intelligence agencies aren't running around discussing their methods publicly, most intelligence agency work doesn't result in public criminal charges.

They’ve built a platform and sales pipeline optimized for selling data consulting into highly bureaucratic tech hostile orgs with data privacy concerns. All these factors apply equally to public health programs and the military, so it’s no surprise that they see success in both areas.

To be fair to GitHub, "GitHub" Actions is just Azure DevOps Pipelines wearing a mask. Which I think explains a lot about it's quality as a feature. It was brought in as a rushed copy-paste of the existing Azure DevOps feature very quickly post acquisition.

I have to regularly use Azure DevOps and the whole platform is painful, and now is rotting on the vine. I hear there is internal strife at Microsoft between Azure DevOps and GitHub products.

Devcontainers (I know it's not a full VM, but it's most prominent version of this "isolated development environment" concept) wouldn't fully protect you against this. Github credentials are automatically pulled into the container. If you are using other cloud services that need to be accessed within the container, this cred stealer will grab their creds too.

It would limit the blast radius, which at least is an improvement.

These malicious packages are being caught by the authors, and by automated package security scanners, not just by end users. npm should start setting this 7 day cooldown as default.

I’d honestly expect to see regulatory intervention if they tried this.

An eventual goal is likely to allow interacting with the LLM directly via audio tokens in input/output skipping tts and stt completely.

This really is a great blogpost, I'm really glad you shared it.

Yeah there is some degree of awkwardness created by the interaction, but I think it’s less about needing specific libraries to map well and more about getting a good understanding of what the interop rules are, and what the shape of the underlying generated output actually looks like.

C# interoperability loosens guarantees (particularly immutability) that F# code normally relies on. There are surprising limits that come up in generics because of how they map to C#.

Fable is great but it has a surprising number of these hidden behaviour changes that are really hard to detect when writing code against it.

My understanding is the uutils development process involved extensive testing against the behaviour of the original utilities, including preserving bugs.

I think there is also the added challenge that ARM macs are a moving target, and Apple has less than no desire to provide any kind of stability or support for Asahi Linux. Unlike the PC space where laptop manufacturers have to maintain broad compatibility over time, Apple will make future changes that are really awkward for Asahi and will not care one bit because they can do the compat work on their own software.

Model output reflects on your input, and the effect is self reinforcing over the course of a whole conversation. Color you add around a problem influences the model behavior.

A "dumber"/vague framing will get a less insightful solution, or possibly no solution at all.

I don't even necessarily think this is a critical flaw - in general it's just the model tuning it's responses to your style of prompt. People utilize LLMs for all kinds of different tasks, and the "modes of thought" for responding to an Erdos problem versus software engineering versus a more human/soft skills topic are all very different. I think the "prompt sensitivity" issue is just coming bundled along with this general behavior.

Does non-artificial intelligence have clean instruction/data separation?

I found the Copilot harness generally more buggy/disfunctional. After seeing a "long" agent response get dropped (still counts against usage of course) too many times I gave up on the product.

It doesn't matter how competent the actual model is, or how long it's able to operate independently, if the harness can't handle it and drops responses. Made me think are they even using their own harness?

At least Anthropic is obviously dogfooding on Claude Code which keeps it mostly functional.

Is Composer 2 a bad model because Cursor are bad at training models, or because they are compute constrained? This deal will provide the answer to that question.

I think it also represents a bet that in some sense Cursor's model capabilities are resource limited rather than talent limited. If that's true, $60B will end up being a bargain. If not true, well it's an expensive lesson but that's the nature of things.

Problem is basically, that if the option works out (Cursor truly has the talent to train a frontier model on SpaceX's infrastructure, and were simply lacking the infra before) the fair price would be way way more than $60B.

OpenAI tried to acquire Windsurf last year for $3B and couldn't.

It's true that Cursor doesn't have their own frontier models, but they are training their own models. They just aren't at frontier level yet. The $60B/$10B deal looks like a bet that this is a capital/GPU constraint rather than a capability one.

The problem is, the ways that it's great are hard to put into words and explain for someone unfamiliar with the game. At least, not in a succinct way. At the most basic level it's just block pushing puzzles ("Sokoban").

Trying to copy is a bit harsh, but he has credited it as an inspiration for what's possible within the Sokoban format. I believe he was a playtester for SSR before it released, mid The Witness development.

If you can correctly add all the required explicit bounds checks in C what do you need Fil-C for?