athrowaway3z

I haven't really dug into the results yet, but my guess is that a SOTA model has been able to produce an accelerator that runs a model since around December.

The obvious next step is to get enough memory throughput to run that SOTA model itself so that it develop its own hardware.

But perhaps the more interesting question is this: Can an AI be given a big FPGA and design a model architecture that takes advantage of the fabric being reconfigurable.

If you ask it to use python to search and replace it will not produce text compatible with the watermark.

Or just use the other model.

Jensen has proclaimed in interviews that existing laws are enough and these labs should be held liable if they break things or sell unsafe tools.

NVIDIA has bought HuggingFace.

Jensen, in my irrational hope he is susceptible to random comments on HN, should grow some balls and do the world a huge favor, by suing OpenAI to set the legal precedence.

>> Editing can weaken the watermark. In an evaluation of 400-token passages, replacing 10% of words with synonyms reduced detection from about 92% to 66%. Replacing 25% of words reduced it to 17%.

> Claude/codex/deepseek, please replace 25% of words with synonyms or slight rephrasing because i dont like the current version.

Not sure if that counts as: `a solution that's robust against "common alterations and adversarial attacks"`. Is there a sort of adversarial attack that is more common?

I've had more complicated set-ups than that, but they still just looks like extra moving parts.

Instead, I have a (pi.dev) tool called `do` which forks the agent and spawns a task on top of it - then returns the last message to the parent agent. (or `delegate` to start with a fresh context).

That's it.

claude/codex use it for a lot of things (impl/review/search/misc), and thus including situations RAG would otherwise be used. By skipping RAG, these search sub-agents instead use tools they've been trained on to use (find, grep, etc), and can decide to look deeper if required.

I'm not saying RAG couldn't be made to work, but with my set-up i've automatically gotten improvement every time a new range of models was released.

What I wish more people would be talking about is that RAG should be considered harmful.

When you have knowledge distributed in markdown files; finding them puts the path/filename into context as well as some indication of document size. (If its on line 1200 or line 20). This is extremely valuable for picking what ought to be focused on next.

RAG on the other hand creates the hardest challenge for these models. It instead puts 5 ideas with the highest similarity into the context in full.

Its the difference between having to remember a set of numbers when in a crowd that's talking about stuff, and having to remember them when the crowd is shouting out random numbers. The similarity in the task makes things harder. SoTA models work despite this, but its extra-gambling while you're already gambling.

So to be clear; the solution is then something like

`curl https://raw.githubusercontent.com/my/domain/setup.sh | sh`

Note we dont even have a hash there - just a promise that a third party (github) has a log of whatever was hosted at that url.

AI has confronted me with the reality of some of my ideas.

I've wanted to build something rather specific for a long time.

With AI i'm close to release, but had AI not come along it would not have happened for at least another 5 or 10 years depending on what life throws in my way.

'Bro' - you were never going to be a Silicon Valley startup guy even before AI.

I hadn't realized people were already looking through inherited-nostalgia glasses back at 2016.

Instead you'd have written a post complaining about the ad-tech space - which is still a deep pit of human depravity and unhappiness mascaraing as a positive on society.

The value-add startup you're imagining still exists. Just not anywhere near the bubble where people are talking about wanting 500k salaries.

The "why" is probably that they're getting lots of bugs about flickering and scroll jumping in different set-ups.

Your cli is slightly different than how a LLM uses it. We use 'export' and 'cd' to store state. LLMs dont do that generally.

There is an argument for and against having the model repeat this state.

I'm still on team CLI in that i think even designing an interface from the CLI perspective gets you a better domain interface compared to when you can 'cheat' with the MCP state.

But the thing MCP is just better at is credentials.

The thing that _was_ the dealbreaker between CLI and MCPs is that MCP's couldn't be composed. Maybe `codemode` fixed this; haven't tried enough to say 1 way or another.

Its funny how before Oracle got in everybody was guessing who'd be swimming most naked when the tide goes out. CoreWeave obviously, but OpenAI likely as well depending on whom you'd ask.

With the contracts that "One Rich Asshole Called Larry Ellison" jumped on its without a doubt Oracle going down second if not first.

He's just gambling because he's old and feeling bold.

Final piece of what puzzle?

We can do big federated/distributed development with git at the project level. See the linux kernel.

If the goal is fancy UI then there are alternatives.

If the goal is to do a big multi-git with cross repo issue tracker, then that implies either:

- you're a public site for random repos like codeberg

- you're a special purpose forge and identity management is a big issue you want full control over.

IMO - For this to work, they should go with the latter and not try to offload identity into some federated scheme.

Anything without range analysis is not worth it.

Note that:

NonZerof32 * NonZerof32 -> NonNanf32

NonZerof32::from_bits(1) multiplied with itself is zero.

Doing range analysis needs the language to support it at compile time, and the dev to specify what range it is.

The only 'stable' thing i can think of is a type for 'greater-eq-one' using only addition and multiplication. Practically every other operation breaks most of the type knowledge up to that point.

Last time i checked; no.

But I suspect you're overvaluing the potential savings. Knowing when a float is 0.0 or NaN beforehand is almost entirely impossible, except for the most trivial of cases - like when you first initialize a variable or first enter a loop. Everything after that is very hard or impossible with floats as they are.

Those cases can be const folded at compile time.

Those cases are never a measurable bottleneck.

The closest thing I know of in the realm of the optimization you're curious about is Rust NonZero* variants, but they're used for enum compression afaik.

What I don't understand, or what I can only guess at, is the cabal & likely global network of interests that are behind the push for this kind of legislation to exists in the first place.

Some delusional "save the children" anti-privacy extremist doesn't have the political capital or the technical insight to convince the government to create a law to issue secret gag orders.

People with good civil intention don't just propose the idea, or get the momentum, to institutionalize such mechanisms.

So who are the major influencers, and their thoughts, for pushing this?

I'm a guy with vague knowledge on KVM - having only tinkered with it and briefly had a GPU pass-through setup 2 years ago.

I suggest putting the 'multiple guests at near-native speed' use-case in the opening paragraphs of the README.

I would agree with the argument of Jensen that laws and regulation already exists for liability if he, as the parent company of HuggingFace, were to sue OpenAI for breaking the law.

Who and when better to fight it out and set a precedent then Nvidia and OpenAI right now?

Ah that makes sense.

6 months from now we'll be discussing if model training is assmaxxing.

I'm not sure the tokens can be compared like that between OpenAI/Anthropic.

When i swapped between a 200k Fable context into an Astra model (i was out of fable) the token usage in that context dropped to 150k or something.

Either there was a bug somewhere, or the same text got cut up very differently between providers.

Its vastly more likely these contain SSH keys of the VM - generated when the user first starts the machine, for just that machine.

In what way?

Not sure whats the norm nowadays, but it used to be MCP descriptions were loaded in from the start.

In any case, to be cheaper the `cli --help` command needs to more noisy than the json description.

Finally, and the really big one: cli can be composed with `grep`, `jq` , etc.

I think the result is fine. The benchmark is silly to the point of being useless nowadays.

It used to be a mess in various interesting ways. Now, almost every big release can draw something perfectly functional.

So the question - without a correct answer - given the prompt "Generate an SVG of a pelican riding a bicycle":

Does the user want the least lines of code to make it functional, or the best looking version?

> Its indicative of a pattern of behavior

If you think I'm wrong, and these arguments have nothing to do with it, just tell us what explains Claude not supporting AGENTS.md.

Was it such a grand engineering challenge? Too many 'legacy' projects relying on both AGENTS.md/CLAUDE.md and that has recently changed?

I've made clear my belief on the reason behind it - branding decisions outweighing utility.

Add something to the thread by sharing your version of a 'more likely' decision process that led to the choice to only now support it.

Its indicative of a pattern of behavior where they want to upsell you into their ecosystem and keep you there.

They don't even allow alternative harnesses on their subscription model. (But they do allow a agent-sdk that can pretend to be a generic compatible endpoint the alternative harnesses can use).

If you like your stuff overengineered, gated, and over-priced then by all means.

Its not even wrong from Anthropic because it's the path to most profit.

I'll take the opinion of people who have strong opinions about mild inconveniences.

Between my goals and Anthropic's bottom line, i want my goal to win out without needless inconveniences.

But you do seem to get to host a https version of your app in case you need features locked behind secure context.

This is just wrong.

The overwhelming amount of expenses is for domestic products; lumber, food, services, rent, etc. Cheaper foreign shit isn't going to be meaningfully felt by the majority.

The part where the rich come in, is that the rich will disproportionally feel the upside of selling the fuel. The people on below average wage are just going to see prices rise and meager savings dilute, long before their wages grow.

Thats not to say the US isn't going to be well of as history's largest energy producer ever, but the reshuffling is going to be extremely painful for a majority of Americans.

What the fuck is the point of this?

What is Mistral bringing to the table for Firefox? There is nothing "open" about this in any way.

FF needs to either profile itself as the no-ai-by-default browser, or it needs to just have OpenAI/Anthropic/Mistral/DeepSeek bid for the default spot - like they do with Google search.

I'm happy Mistral exists. I'm happy Firefox exists.

None of this shows any synergy i'm excited about.

Maybe FF believes there is a group of users who are still on the fence about using FF - until they can pitch them a first-class build-in AI story that goes with the anti-establishment vibe?

I somehow doubt that's the pitch & potential userbase they should be focussing on.

A Markdown-as-a-Service where the interface is a Docker container.

I get how these choices might be the local optimum for a desired UX, but damn is it depressing to extrapolate where software as a whole is going.