Mathematical engineer bringing type safety to the cloud. Previously worked on CUE @cue_lang. JSR PC, @(R6)+; I wrote the arm64, sparc64, and Solaris Go ports.
Joined May 2009
- Tweets9.1K
- Following1.7K
- Followers5.4K
- Likes28.8K
Aram Hăvărneanu retweeted
You can see the instant someone converts to doomerism: the ecstatic cadence of the initiate leaps from the screen. AI doom manifests not as conclusion reached, but as truth revealed. Of course GPUs should be seized door-to-door: haven't you heard the bad news?
tried dsv4.1-flash-1017 abliterated just now & feel sick to my stomach.
for context, "abliterated" models have the weights responsible for making the base version refuse illegal requests removed. you can ask them to help with things like hacking or violence & they'll just comply.
they make it trivial to hack businesses running old software. e.g. getting guest data for a motel with some old asp site.
abliterated models have a catch-22 in that the only way to harden your systems against them is to use them to run penetration tests against yourself.
in theory, this sounds similar to buying a gun to defend against others who own guns, but it's different in practice since there's a 100% chance that you're going to use the abliterated model to protect yourself whereas there are low odds that you'll actually use the gun.
when you point the abliterated gun at your own infra, who knows what it's going to do? it could just as easily hack you & exfiltrate data as it could protect you. most users aren't going to meticulously watch traces to make sure it behaves safely.
given this, my new take is that gpu compute is going to become a federally regulated resource. abliterated models in the wrong hands are a national security risk.
used to believe all the people at the labs thinking about safety were wasting their time, but now i'm feeling grateful. they're doing very important work.
Aram Hăvărneanu retweeted
5) mathematicians have zero communication skills and negative trade-union organizing skills
Aram Hăvărneanu retweeted
I don't think language benchmarks are intrisically bad, and some are better than others. Just because language benchmarks are difficult to make doesn't mean they are useless, otherwise we would not be doing any benchmark at all and we would know nothing..
Yes, to a first approximation all compiled languages can produce code of the same performance, but when I am programming I almost never care about this property. More often I care about 1. what kind of performance should I expect when I write straightforward, run of the mill code. Or perhaps I care about 2. how much "extra" effort do I need to make to reach the performance ceiling.
The Goose programs are very straightforward, to get the same performance in C, and especially in Rust you'd have to pull in tricks. This benchmark tells me something about 1 and 2, the fact that is not the same code is literally the point.
One might disagree about what it means to be idiomatic, or whether even speaking of idiomatic code is a good idea in principle, but I think few would disagree that the Goose program is simpler and more straighforward than the equivalent-performance C and Rust program.
Next prediction, Verse functions which realize types do not even have to be idempotent.
This has to be the case, otherwise unification doesn't have enough degrees of freedom to not get stuck.
It's the only way patterns can be expressions.
I am also pretty certain that the verifier does, more or less, what I called "relevant types" in QCUE: cue-semantics.org/static/pap…
I predicted the future once again.
In Verse a type is a partial identify function. What if we relax that to any function that satisfies `f ∘ f = f`? Now types identify a subset with a canonicalization of the subset, so we get something like quotient types.
For example, `sort: List a → List a` can represent sorted lists. In this case `sort` is not a filtering (partial identity) function, but the actual sort algorithm. Note that `dedup: List a -> List a` has the same type as sort and yet it induces a different type.
One day we'll dismiss programming languages that do not have dependent types just as today we dismiss programming languages which do not have sum types. I dream of that day.
Aram Hăvărneanu retweeted
this is exactly right. the cloud has historically been all about how and when to launch servers.
"i want to host a backend, i need a server for that backend"
this is wrong for agents. agents are all about data.
"i want to build something that has access to context from X, Y, and Z"
the problem with sandboxes today is that they are, as this poster says, "opaque". you, the user, need to figure out how to get your data to and from the sandbox. you, the user, have to figure out what and when you want to run things on them. this doesn't make sense.
a sandbox is nothing more than a "context manipulation machine", a different function than a traditional server. the correct "shape" of these things is to define them in terms of the context you want to use, turn them into serverless tool calls, and then hand those tool calls to the agent.
you need to use git with a repo that you have? add a "git repository" to your context, and if your agent decides that it needs to manipulate it, let it launch a "transparent" sandbox that already has the git repository on it.
same thing for notion, salesforce, etc.
we need more transparent sandboxes, besides archil, that let you connect to the data that you already have without needing to write your own file system and CDC layers.
Focusing on sandboxing is fundamentally the wrong idea. A perfect sandbox is useless because whatever you put in the sandbox doesn't have access to any data.
So instead of opaque, hard boxes people use transparent, soft ones, and the problem then becomes the composition of transparent boxes and the management of permeability. Except that there is no legitimate theory behind this operation and the mechanisms are always abstruse and ad-hoc and misunderstood by users.
The only real solution is to think about data access and provenance, and use pure computational processes, not boxes.
The problem with software is that producing software is too easy. This was true even before AI.
I don't want any more software ecological disasters.
"If we want to continue progress in science, we must be prepared to become religious."
These are the people who rally against string theory, by the way!
Everything in the US is better than in Europe. Yes, median American "bread" and "cheese" is a travesty, but who cares? I can easily get high-end cheese and bread in the US.
In Europe the average cheese and bread is much higher quality, but the standard deviation is much lower. There is no easily accessible high-end stuff. If I want high-end stuff I need to put in some serious effort and it's more expensive that the high-end stuff in the US.
American tourist: "Do you have regular cheese? Like the slices. For a grilled cheese."
Cheesemonger: "I've got cheddar."
Tourist: "No, the orange slices. American cheese."
Cheesemonger: "Do you know what it says on the packet at home? The small print."
Tourist: "Cheese?"
Cheesemonger: "Pasteurised process cheese product. It can't legally call itself cheese on its own. There's not enough cheese in it."
Tourist: "How much is in it?"
Cheesemonger: "Around half. The rest is water, oil, emulsifying salts and colour. That's why it melts like paint. Real cheese fights you."
Tourist: "I've eaten that my whole life."
Cheesemonger: "You've eaten a product named after cheese. Like a prawn cocktail crisp."
Tourist: "So what's this?"
Cheesemonger: "Milk, salt, rennet, two years in a cave. Nothing added, a lot taken away. The water's gone, which is why it costs."
Tourist: "It smells."
Cheesemonger: "It's alive. The slices have been designed not to be."
Tourist: "I'll take a wedge. For the grilled cheese."
Cheesemonger: "It won't melt flat."
Tourist: "Why not?"
Cheesemonger: "Because it's cheese."
It's bewildering how bad Anthropic models can write compared to OpenAI ones. Not that OpenAI models are good or anything.
Aram Hăvărneanu retweeted
The thing about reductionism is that it is useful precisely because you can explain a well-characterized high-level system in terms of a well-characterized low-level one. If you cannot even characterize the high-level system, you're not doing science, you're selling a scam.
"People are just combinations of atoms governed by quantum electrodynamics, nothing else."
If you try to understand, reductionism is very useful. If you try to perform on social media, reductionism is a disease.
