For thorny problems I let the agent give me a simplified flow-chart in mermaid syntax. LLM's brain-farts are easily visible then. I correct the flow-chart "Ah, you're right!" and then let it translate it to code. Works wonders.
Well I've seen worse bookkeepers. "You know, you approved of the budget, but where are our customers payments in the balance sheets? We can't find them!" - "Uhm..."
I believe it. I experienced this once, as I tried to have everything owned. Now I just clone around as if there's no tomorrow and tell myself I'll optimize later.
I'm kind of impressed by the speed of it. I told it to write a MQTT topic pattern matcher based on a Trie and it spat out something reasonable on first try. It hat a few compilation issues though, but fair enough.
I remember vaguely that this was observed when training GPT-3 (probably?) as well. Just trained on and on, and the error went up and then down again. Like a phase transition in the model.
I started my most ambitious project in February. A few years ago I wouldn't have even dreamed of ever starting let alone finishing it, but now I have a Claude Code Max Pro subscription and it goes forward in a steady pace. I expect the first version of it to be finished within the year. Even it's written mostly by AI, it's still a lot of work to get it to do the right thing, but I'm getting better at it.
I always dreamed of a tool which would know the intent, semantic and constraints of all inputs and outputs of any piece of code and thus could combine these code pieces automatically. It was always a fuzzy idea in my head, but this piece now made it a bit more clear. While LLMs could generate those adapters between distinct pieces automatically, it's a expensive (latency, tokens) process. Having a system with which not only to type the variables, but also to type the types (intents, semantic meaning, etc.) would be helpful but likely not sufficient. There has been so much work on ontologies, semantic networks, logical inference, etc. but all of it is spread all over the place. I'd like to have something like this integrated into a programming language and see what it feels like.
Interesting that the film material had better color retention than the old projectors. I never thought about it before and assumed that the washed out colors of old 16mm projections came from bad recordings.
Some films were recorded with a lower frame rate due to some reason. As for variable and standstill, I think this would only be needed when investigating a recording in detail. Not sure if they really need it. Perhaps they just want it to cover more use cases.
I recently investigated some problematic behaviour of both Opus 4 and Sonnet 4. When tasked to develop something more complicated (broker fed task management system, staggered execution scheduler) they would inevitably produce thousands of lines of over engineered, unmaintainable garbage. When Opus was then tasked to simplify it it boiled it down to 300 lines in one shot. The result was brilliant. This happened twice.
Moral of the story: I found out that I didn't constrain them enough. I now insist that they keep the core logic to a certain size (e.g. 300 lines) and not produce code objects for each concept but rather "fold them into the code".
I mostly use it to understand and fix my bugs/Rust compilation problems that I can't be bothered to fix. 95% happy with the results so far. For coding I use Claude in chat though, as my thoughts are mostly not clear enough at the start to finish a component to my liking. Fixing bugs is easier, though I had to tell it to "not remove features" sometimes. Feature gone, bug gone. ;)
Claude code now automatically integrates into my ide for diff preview. It's not sugar, but it's very low friction, even from the cli.
From my experience operating it I reckon it does the following: after you fill it and start, it starts heating and measures the time it takes to reach 100°C. This allows it (in the rice program) to figure out the amount of mass present and adjust the simmering time under pressure accordingly.
Using the same trick (in alll programs) it figures out overfilling and aborts for safety reasons.
Edit: I think the only sensor needed is a temperature sensor.
I've been to Café Central. There was a waiter standing outside telling you only to come in whenever a seat has been vacated. Once in it didn't feel crammed, and it was quite the sight. Don't eat "Sacher Torte", it's quite bland for today's standards. The Kaiserschmarrn was fantastic though.
Parents thinking was probably: If you can achieve similar results with a fraction of memory/compute usage then capability at the same hardware level will increase even more.
ECC RAM would have probably helped, but that got axed in consumer CPUs probably due to financial optimizations. They wanted to have ECC as an upsell feature for 'server grade' products.
> Not because of conscious fear, simply by default. If Navalny became Russian president - the next day 80% of Russians would be completely persuaded they were always against the war.
This seems similar to how the people in Germany reacted after the Nazis lost WW2.
This technology is useful for small production volumues only though. It's too slow to produce chips at 5nm at high volumes. If it's also that much cheaper such that you can just throw money at it to produce in parallel, then it may work, but it I haven't seen such comparisons yet.
To install Windows with only a local account, you have to disable Internet access (unplug cable or disable WiFi), then that option magically appears. Once done, reactivate access and continue. For most other cruft like games and such there exists a intimidating command line one-liner which gets rid of it.
Regarding gasoline: the question is how much of that energy is actually converted into motion instead of heat. With electric mirrors it's more than 90%, with gasoline it's more like 30-40%.