I dropped Pascal programs in there, because it was the language I knew best, good times! It was really slow, too, because the bundled runtime had a massive startup time.
I think the approach here is to code these basic features by hand. What's preventing the author from building what they need on top of their basic stack?
Speaking as someone who hand-rolled their server stack as a C++ framework for years, there is nothing in there that can't be done relatively easily. Some things certainly get easier even, the actual change in paradigm here is not that it's a hand written web app server in C, it's that it runs in its own server container which you have to manage. That's the actual tradeoff. You can't just deploy to Vercel.
- start with zero skills
- add a skill if you encounter behavior that you want to ward against or if
you want to associate a meaningful phrase with a certain method of doing things
- NEVER copy a skill from someone, do not clone skills repos, do not let LLMs
write their own skills
- occasionally revise or delete a skill, less is more
This looks very promising. It would be cool if support for Bonsai-style models would land in mainline MLX soon, looking forward to trying it out.
It seems PrismML has implemented a better version of an idea I had a while back: what if we had a 1-bit model where the scale of the weight is determined by its position. The model would have to be trained from the ground up for this though, which is why I never tried it. The interleaved scale factor approach of Bonsai is a much more flexible approach at almost the same cost.
I know this is an outsider position, but I always felt that the AoC leaderboard was a mistake. Very few people had the time, the commitment, and the capability of making it on there in a meaningful fashion, and it put an emphasis on something that didn't match the vibe of the event at all. If speedrunning the problem solving was the point, then why package every episode into an enjoyable little story?
This also ties into the comments that AoC has become moot or was "ruined by LLMs". If you enjoy solving the problems, nothing should have changed for you. What's the difference if a given problem was already solved by an LLM, or a group of IQ 200 superhumans from MIT for that matter?
As time marches on, there will eventually be absolutely nothing left where an unaugmented human outperforms a machine. That doesn't mean you have to stop enjoying things. In a few years at most, all programming will be purely recreational.
> In a few years at most, all programming will be purely recreational.
That's a bold prediction given how much LLMs suck at programming today (and haven't really improved, either). I'm willing to believe that we will someday invent an AI that can program better than humans. I don't believe it'll be within a few years, because the current architecture shows no signs that it'll ever be able to get the job done.
Exactly. And as implied by the term techno-feudalism, the owners are okay with a greatly reduced economy, and in some cases a severe reduction in quality of life overall, as long as they end up ruling over what's left.
If the candidates are legally bound to follow through on policy then it's slightly better yes but there's still a lot of wiggle-room in terms of implementation. Lobbyists often find ways to sneak in convenient loopholes inside regulations. Policy-based voting won't solve that because the voter isn't going to have time to read the full law; it might be many pages. What will happen is people who don't work and have more time on their hands will be more active politically and will get better representation. I mean, it'd definitely be an improvement though.
I had to try this gem, it's my new benchmark! o4-mini-high also fails spectacularly, even after repeated feedback. However, 4.5 (the impractibly large demo model) gets it right:
It’s a nonsense twist on the classic lateral thinking puzzle:
The original puzzle goes: “A father and son are in a car accident. The father dies, and the son is rushed to the ER. The surgeon says, ‘I can’t operate on him—he’s my son.’” The intended answer to that puzzle challenges the listener to recognize the surgeon is the child’s mother, confronting implicit gender biases.
Your version humorously mangles it by swapping roles (“son and his man”) and objectifying the victims as cars, creating a logical absurdity. The sentence “I can’t operate on this car, he’s my surgeon!” doesn’t parse logically, indicating it’s a playful distortion rather than a coherent scenario
Yes, skip would be good, but I'd also advocate an option like "I reject the premise of the question".
In legal contexts yes-or-no answers can work because the case can in theory be boiled down to guilty or not. If there is any flaw with the case, the answer should be not guilty.
But let's take the "do I have a moral duty to..." questions used as examples here for contrast. I'd argue you never had a moral duty to attend your sibling's wedding to begin with. But because the question was asked with a weird modifier like "even if it's their 3rd wedding", any answer you give will be inadequate and will just serve to reinforce the flawed premise. Skipping is not enough in my opinion, because even if communicated to the question asker, it doesn't make it clear whether there as an issue on the answerer's side ("I don't know" / "don't feel qualified") or with the question itself.
What you're describing sounds like the curvature or topology of space would be non-flat. AFAIK this hasn't been completely ruled out, but so far every piece of evidence suggests the universe is flat over vast distances.
Intuitively I'd say if there was curvature or topological irregularities at the furthest distances we can observe, there wouldn't be a consistent redshift observed on far objects because some of them would be coming towards us instead of pulling away.
reply