Small change in comparison, but today Fable created and benchmarked an RTree which was 100x faster to populate and 5x faster to query, compared to a previous attempt with Opus 4.6 a couple of months ago. That took about two coffees.
I think it's important to remember that as impressive Opus & co are, they're standing on the shoulders of giants, i.e. the engineers, academics and companies who have cooperated to design the incredible programming languages and machines we have today.
I wonder if this is this just the law of diminishing returns at play?
My thinking being, it's been a few months since I thought the code generation machine was the problem, rather than my interactions with the machine. A month is a long time in AI.
What I mean is, these things are about as smart as they need to be already for the average SWE. I don't think this is true for those solving the really big questions like curing cancer.
Also on a macbook, it took me a while to realise pinch to zoom doesn't do anything except zoom the page, pan is click and drag, and scroll is zoom. Cool project except for the controls, surely scroll should be to scroll up and down?
Looks like Apple are making an effort to get agents to create decent code with Xcode skills, which someone has extracted here [1] but there is near zero mention of accessibility. Worth a read though I think, I learned a few things.
Either way it's a cop-out. At least make the effort to have a cleanly built, well-understood system. Sure there will always be poor practices built into a system, but the number of places this happens can be mitigated.
Vim is a bad example to use here. I certainly have no hacker-vibe pretensions, I use Xcode almost exclusively, but with the Vim key bindings which have been integrated very well. The productivity gains with Vim commands are impressive.
I do hope Anthropic are better with IP, and I think they may be. Given Dario Amodei hasn't been sued by OpenAI while building Anthropic this seems likely.
I think Amodei may actually be quite a good human, despite my trust in big tech being at an all-time low.
I loved his Algorithms course on Coursera which I did during the brief moment Stanford MOOCs were all free. So useful for a non-CS grad doing any kind of algorithmic programming.
In the same sense that a blockchain can be forked by using software that only accepts certain types of block, is it possible to fork the WWW in a similar manner? e.g. with changes that neuter the ad-mongers.
For example coming up with a way to get rid of these god awful cookies. Maybe ad-monger sites could be allowed in the same way an insecure connection is allowed behind a series of warnings?
Small change in comparison, but today Fable created and benchmarked an RTree which was 100x faster to populate and 5x faster to query, compared to a previous attempt with Opus 4.6 a couple of months ago. That took about two coffees.
I think it's important to remember that as impressive Opus & co are, they're standing on the shoulders of giants, i.e. the engineers, academics and companies who have cooperated to design the incredible programming languages and machines we have today.