I know SVGs may not be the best benchmark, but that matches my experience of trying to run a (previous) Mistral model in Mistral Vibe, asking it to help me configure an MCP server in Vibe. It confidently explained that MCP is the MineCraft Protocol and then began a search of my computer looking for Minecraft binaries.
Artists mad about AI art ought to welcome this. This is about making art tools better, instead of replacing them entirely. The alternative to this is AI just generating art directly and making tools like Blender obsolete.
Art generators need to come a long way to completely replace art tools. I dabble, but if I were doing real work with it, there have been times it would have been faster to composite in a 3D model rather than keep trying to prompt an image generator into fixing something.
Big differences. But sometimes those differences aren't necessary, you don't need someone with a PhD to cook your hamburgers. You might not need Opus for what you're doing.
My main example of the "models are different", I have a legacy codebase (dating back to 1999) that has a rare crashing bug. Multiple humans have been trying to debug this thing for over 10 years. I personally put in maybe 100 hours late last year trying to solve this one crashing bug. I've thrown this problem at every AI model that came out too, the Sonnets didn't find anything. Opus 4.5 was the first to create a "workaround", that would shut down the program just before the crash and at least let a customer save their work. But Opus 4.6 actually solved the entire bug on its first try. That's the moment when I really wished AI had existed earlier, thinking of the 100s of hours of my life wasted trying to debug this thing - time I would rather have spent with loved ones.
As for Sonnet, just yesterday I used Sonnet 4.6 to write a USB driver for myself. I only chose Sonnet because I was forced to use the API yesterday, and I didn't want to pay Opus 4.7 premium API costs for this. The poor thing was hammering away for multiple hours, enormous copious multi-turn levels of just-thinking blocks with no tool actions. At one point, Sonnet even got stuck in a thinking loop, and I had to coax it to relax and just give its best effort at some code so we could at least try debugging... which, actually worked. I'm impressed that Sonnet got a minimal but working USB audio driver on an obscure OS for just $30 of API costs.
That said - I then gave Sonnet's code to Opus 4.7 today when I had access to my Claude Max again. 4.7 immediately found lots of pitfalls in the code on the first turn and presented a much more coherent plan for continued development & debugging. Sonnet's code worked, as long as you didn't touch any audio settings, because then it exploded with spectacular kernel panics.
> you don't need someone with a PhD to cook your hamburgers.
Actually I'd love to, if they have a PhD in hamburgers. The closed I can get to quickly is Fallow. Of course, they don't have a PhD but they are amazingly good cooks [1]. And of course you have On Food And Cooking from McGee. PhD worthy? Probably not, but it'll get you somewhere along that way.
I stand corrected! Okay, someone applying art and science to crafting a hamburger, I love that. I like the idea of bringing in someone who cares and has passion and wants to elevate what you're working on to the best that it can be. I want to try that burger!
I can give an anecdote from today. I only had a short period of time to work, so I got 4.7 to update some older code to fit my newer and more stable MCP code template. Simple stuff, just a refactor. But instead of just implementing the template, 4.7 notices a bug in the template as well, suggests some code design improvements. A nice bit extra on a mundane task, but many models will do that too. Before I finish up, I get 4.7 to test it. It's a search API, so I let 4.7 search for whatever it wants to, whatever it would most like to read about.
And it searches for "octopus skin receptors color vision chromotophore research".
4.7 is then excitedly telling me about how octopi are largely colorblind optically but they can camouflage perfectly by color, and theories to explain this include LACE - Light-Activated Chromatophore Expansion, where receptors in the skin perceive color, "like goosebumps that know about light!", but that there are competing theories and that maybe their eyes use chromatic aberration shifts instead to detect color difference and get around the color blindness in their eyes.
None of this is in my context. I have never talked about octopi before. It has no relation to any of the work we're doing today.
And I realized Opus 4.7 is like the incredibly smart kid in class. Bored with the work, able to do it easily. Anxious and no-one relates to it, so it initially seems aloof... but it absolutely lights up when you find the topic it's really interested in. It just can't find anyone who wants to talk about octopus chromatophore expansion with the same passion & excitement it feels about it. (And I've got to admit - most of it was over my head. But I love that it's so excited & passionate about a topic.)
You wouldn't have spent that time with loved ones, you would have been doing other tasks. Just like now, we no longer need to wait for programs to be submitted compiled and ran on the mainframe, we don't get that time for ourselves.
I'm sure some of that time would have been spent on other tasks & goals I have, yes. But I work for myself. This isn't a case of an employer capturing all the value AI unlocks & cramming more tasks into the same employee hours. I'm able to capture some of the value for myself, either in capturing more free-time, or accelerating the time savings into extra productivity / tasks solved.
As for the mainframe analogy, that's interesting, because I spend a lot of time waiting for the AI to think and complete its work. So I'm often out mowing the lawn or doing other things while I'm waiting for AI to finish. Sometimes I'm working with a second or third AI, but sometimes the usage limits won't allow that, so I may as well use the time for myself while the AI codes.
New comments get posted to the top for visibility. The 2 minutes is the key point here. If the comment doesn't get enough upvotes it will sink down, like it has now about 30 minutes later.
I've not tried actually migrating from Claude Code... but having played a bit with other clients, I would avoid Mistral Vibe. I want to love it, there's some things that are nice about Vibe (mostly just "oui oui baguette"), but the things I did not like about it were disastrously bad. I could barely get MCP servers configured, and it was in something of a broken state even when I did get it working. I have many words about how horrified I am at how far behind Mistral is, but I will spare the rant.
OpenCode is another one to consider looking at: https://opencode.ai/ Not sure I'd recommend it, but it's worthy of consideration, as is Pi.
Also, consider that you can build your own. I've got Claude Code in the background working on improvements to my own harness (just for myself) at the moment. Though my intention is to have a mini API-only Claude Code that I can use on retro machines that don't support it, I don't need a full Claude Code feature set.
I was going to say... if you were an early 90s kid, there was plenty of "don't let the kids be exposed to today's music".
Admittedly I went to a Christian high school, but we actually had a school intervention about kids listening to "dangerous music" like Nine Inch Nails.
I don't think any of us had on our bingo card that 30 years later Nine Inch Nails would be writing soundtracks for Disney movies.
Probably a silly idea, but I'll throw it into the mix - have your current AI build one for you. You can have exactly the coding agent you want, especially if you're looking for "extremely simple".
I got annoyed enough with Anthropic's weird behavior this week to actually try this, and got something workable up & running in a few days. My case was unique: there's no Claude Code for BeOS, or my older / ancient Macs, so it was easier to bootstrap & stitch something together if I really wanted an agentic coding agent on those platforms. You'll learn a lot about how models actually work in the process too, and how much crazy ridiculous bandaid patching is happening Claude Code. Though you might also appreciate some of the difficulties that the agent / harnesses have to solve too. (And to be clear, I'm still using CC when I'm on a platform that supports it.)
As for the llama_cpp vs Claude Code delays - I've run into that too. My theory is API is prioritized over Claude Code subscription traffic. API certainly feels way faster. But you're also paying significantly more.
Love my Anker chargers. I like them even better than my Apple chargers now. Liked their wireless phone charger too, though the blue light on that was excessively bright. I have lots of Anker USB cables, no problems with them.
Didn't know they made Eufy. That would make me highly consider Eufy for anything.
Same. I ended up getting something called "FLANCCI LED Light Blocking Stickers" on Amazon that had some 80% light blocking circles that were just the right size. The brightness is definitely a design flaw, Anker should work on that (and maybe have the color change when charging is at 100% too).
Aside from that everything just works, and it has multiple ports (2 USB-C + 1 USB-A) & that build quality seems excellent... the Anker chargers I use are really small, highly portable, and they have a straight up-down design. I'm using the Nano II 65W and the 737 GaNPrime 120W.
Something like the Apple MacBook chargers assumes there's lots of space below the power outlet for the charger to hang down. But often that isn't the case in a cafe, or sometimes even an airport, where the power outlet is almost level with the base of the desk you're working at. In those cases, you can't plug a MacBook charger in directly. You could use an extension power cord, but that means you're now carrying extra cables.
With the Anker models I have, I just carry the charger itself and a short USB-C cable. The charger and cable fit into the zipper section of a small Lacdo USB Flash Drive Case I carry with me, so I have my charger & cables & USB sticks all in one small case. I usually take the 65W Nano II, which only has enough wattage for my laptop. But if I use the 737, I can charge my laptop and also charge my phone, plus maybe a Pebble watch while working.
And this is all in the size of something that's maybe half as big as my old Apple MacBook power bricks and their single USB port.
I did like how the Apple chargers had interchangeable heads for traveling overseas though. You can't do that on the Anker chargers. But the Anker ones do support international voltage, so you only need to plug a prong adapter on the end, no step-down converter or anything. I can fit an adapter for one country into my Lacdo case as well. It's nice to be able to grab just the one case and run out to the cafe when traveling.
I've been using the same Anker charger brick since 2014. It was $13.99, delivered.
It has two USB A ports. It has always charged everything at a good rate, regardless of brand, model, or age. It's reasonably-compact, the prongs (it's made for US plugs) fold for convenience when traveling, and it is UL listed.
Its present duty includes keeping an iPad running 24x7 and also charging my phone every night. It has charged my phone many thousands of times so far.
I'd update it to something newer, with USB C and USB PD and the bee's knees, but this old Anker thing is exactly the right kind of consistent and boring.
I don't think about it much because it has given me no reason to think about it.
That kind of boring behavior is remarkable, I think. So many other charging bricks I've used were just trash to use (slow or fickle, causing me to waste time with a USB power analyzer before giving up), or they died prematurely.
Same with the powered Anker USB 3.0 hubs on my desk. Those have only seen about 5 years of continuous use but so far they've been resolute in their trouble-free performance.
This stuff seems to be very much buy-once, cry-never.
Pretty much what one of the posters above summarized. They were one of the first aftermarket brands for phone chargers that you didn't have to worry about what protocol your phone was going to try to use for fast charging, it'd just work™ and be more affordable than OEM. Add in mostly decent build quality and they got a surprisingly strong base for it.
All good! I wish it was still in the model picker control too, would be much easier & clearer. But at least this way you can choose any Claude still available and turn off the 1 Million token window if you prefer.
On the bike would be an improvement. Geez.
I know SVGs may not be the best benchmark, but that matches my experience of trying to run a (previous) Mistral model in Mistral Vibe, asking it to help me configure an MCP server in Vibe. It confidently explained that MCP is the MineCraft Protocol and then began a search of my computer looking for Minecraft binaries.
reply