Compared to what a lot of companies spend on software for chip and electronics design (we're talking about $10k-200k/seat per year), AI coding assistants have a long way to go in cost before companies won't be willing to pay for them. Companies pay a fortune for software when it enables their engineers to be productive.

For my company, I'd honestly pay $4-8k/month for Claude if I had to (it would be painful, and I'd try to get cheaper options to work first). I know some enterprise Claude users are paying that much now since they have to pay for API tokens. I am certain it's at least a 2X productivity booster for our work. Compared to the cost of hiring another developer, it's well worth it.

If they stop subsidising Claude Code for the pro/max users, there will be a lot of people priced out of it, especially the casual developer. But I don't see it going away for commercial use, even with a large price increase.

Nah dude I think it’s worse than that

Old coding is done, as a workflow in teams. It’s the top down executive pressure of being non competitive as a company, and the bottom up pressure of human laziness

Show me people handwriting code à la NASA

And I mean we as coders have been trying to do this workflow for a while, I personally would refuse to code without IntelliJ magic complete

For this workflow, there’s no going back. What’s hard to imagine is AI taking over the other workflows we predict it will; Customer service AI sucks ass for me as a customer, et cetera

I don’t mind bad customer service AI any more, since it’s my claude interacting with it, not me.

> For my company, I'd honestly pay $4-8k/month for Claude if I had to (it would be painful, and I'd try to get cheaper options to work first). I know some enterprise Claude users are paying that much now since they have to pay for API tokens. I am certain it's at least a 2X productivity booster for our work. Compared to the cost of hiring another developer, it's well worth it.

That enterprise cost you're willing to pay is correlated to how much developers will work for. When driving an agent, almost anyone can do it (almost no skills required).

If devs cost $1k/m, enterprises are not going to be willing to pay $4k/m for Claude.

What I am saying is, there's an equilibrium that will be reached; the price of the human driver and the AI worker will approach each other.

Where they stabilise, I still don't know, but I'd be very surprised if, in any field (not just dev), the human gets paid multiples more than the agent they are driving, as the agents get more capable.

You also have a cost of team interaction going with square if team size or sth.

The papers were all in a 'preprint' directory in GitHub. Academic mathematicians seem to think all research should happen in secret until it's completely proven.

But how much more progress would have been made in the last 100 years if they collaborated in real time? Watching the work someone is doing and spotting errors, or making suggestions would be a good thing. Unless the goal is simply to claim credit for a discovery vs the discovery itself.

It's a little strange to write a comment attacking strawman "academics" with the structure of a partisan political attack ad in response to OpenAI making a mistake.

My intention wasn't to attack academics. I was just trying to point out that being wrong is something that happens all the time, especially when something is in-progress (pre-print). How many papers written by people are withdrawn during peer-review, or while in the preprint state?

If someone had a blog about their research on a math problem, and they figure out one day that what they wrote a week before was wrong (retracting it), would that be a bad thing? I don't think it would be bad, but that's not how academics approach things.

If corporations didn't exist and we all sang kumbaya the world would be a better place, but that's not going to happen is it.

What are you on about. The papers are held to scrutiny, this is how it works. If you put something out there, you should have done the due diligence of checking it (which is how these errors were found, with their own weaker model!). Being a smug prick about checking results that were announced as progress is less than worthless.

Um, it's called peer review. Dumping random crap as preprints isn't progress.

Sadly though, you have to do the cost/benefit analysis of the legal process and your likelihood of recovering anything.

I spent $18k in legal fees over a $22k claim in a construction dispute. I won the suit and was awarded legal fees. So I'm owed $40k plus interest. I've collected exactly $0. The last lawyer I spoke to said I need to cut my losses in legal fees at some point because from a practical standpoint, winning damages isn't the same as collecting them. Especially if the defendant isn't local and has few assets.

The main two issues are wildfire and inversion.

As you said, Boise is in a sort of bowl (more of a valley), in the winter, a cold air system can get trapped in the valley by warmer air in the mountain area. This can lead to some pretty bad air (not nearly as bad as wildfires, just not very fresh). To me, the bigger complaint is about not seeing the sun for days. Idaho has a lot of sun throughout the winter, so it's noticeable. It isn't all winter, just a week or so at a time a few times per winted, depending on weather.

But otherwise, Boise's air is quite clean.

What kind of parameters do you end up changing, and how much difference does it make. Perhaps I am missing something and get more tok/sec, but I usually just do a git pull, then rebuild the latest whenever I get a new model.

In the past, I had to play with chat templates for some models to work with agents for tool calling. But I've never had to do anything other than specify the model, and tweaking the context size in some cases.

Things like the number of GPU layers, whether to use flash attention or not, enabling MTP, playing with context size.

What hardware do you run? I have a first-gen mac studio, and I just run cmake and build with no special options. Same thing with llama-server, I just specify the model and use the built-in web UI.

For reference, I get ~26 tok/sec with the new Muse 30B model.

An M1 Max MBP manages roughly 10 tok/sec without the Dflash speculative draft support so that tracks; the M1 Max apparently has trouble actually saturating its memory bandwidth.

I'm left handed and I never really learned a good grip I could write with. So I've always struggled with handwriting. I also have an essential tremor in my left hand. As I have aged, the tremor has gotten worse. So now, I have to grip tight to minimize the tremor.

I've pretty much given up on handwriting, it's embarrassing to either try to write, or have someone else write stuff for me. I ask my wife to fill stuff out for me all the time, but people probably think that's weird when they see it.

That sucks, sorry to hear it. Have you tried seeing if a tablet handwriting recognizer can be trained on your writing?

My guess is design (features/functionality, not code). When you don't have to write every line of code and you can quickly iterate on features, you have a lot of freedom to dial in what you really want out of an app.

To be clear, I didn't mean this as an anti-AI gotcha. They also said:

> We get feature requests, improvements, ideas, feedback

So maybe I misunderstood, but it sounded like the design was external (and based on an existing product to begin with).

Also, my understanding was that "vibe coding" meant more of "make it do X" as opposed to "here's a design for X, implement it."

Sorry, I was trying to be glib by putting "vibe" in quotes, because I think that term is used for people who have zero experience in software engineering, but when used by senior engineers, it's something else. But, I don't know if some populations would still call it vibe coding.

The design is external in that the requirements come from external customers or internal customers (since these applications) are used by both. We're not trying to duplicate or replicate any existing system, but I can't confidently say that I'm not drawing from any prior art.

Hope that makes sense.

As a senior eng, I use "vibe coding" as a http://www.catb.org/jargon/html/H/ha-ha-only-serious.html

Obviously I'm not just shipping whatever it spits out, but the frontier models and harnesses are getting so good that I'm definitely not checking everything it does or all of its reasoning

Makes sense! My mistake for assuming it was the former definition, lesson learned.

Hopefully soon our industry will align on some standard terms. I feel like "AI-assisted", "agentic coding", "vibe coding", etc each have a few different meanings already.

YES!!! Very well said and captures what I was trying to convey.

There's also certain features of an application that most of us engineers know how it works or how to do it, but it is just so painful to do it by hand. Or other features that you've always wanted because you saw another app do it and it was beautifully done and you can just "have" that feature in your app too.

The joy is in seeing the feature come alive, not so much in fighting the computer.

You'll have to take my word for it because I'm not going to disclose my private business or personal financial records.

But, I'm running a small, two person business and we are being paid by a large company to develop a robotics project. We're working at a very fast velocity and have fielded a prototype within a few months. I've been doing this kind of work for years, and we're more productive than larger teams (8-10 developers) I've employed before. Five years ago, I would have needed at least 4 senior engineers to do the work we're doing now, and we're moving faster than we could even then.

And I'm being compensated at a flat rate by a major company so we're making really good money. The customer is happy with the results. Claude easily writes 90% of the code we use.

There's other word for such code and it sounds similar. It's crappy. I was tempted to use Claude couple of days ago to help me with optimization and from 4Gb RAM and 21s it managed to squeeze 1Tb and 41s. When I asked to fixed it AI reduced it to 1Gb and 10s. It just removed most heavy business logic. So yeah. 3 days and this was only total waste of time and cash

How do you know it’s crappy? Have you read the code?

Your comment is strange, he clearly did in this context - that's why he knew what went wrong?

Fwiw, I currently consider code generated by models below opus level borderline unusable. It really is horrendous there.

With opus, it's still not good - but it's a lot better then some of my colleagues write.

Well written code, carefully planned out and thought through is still only possible by either hand writing or holding the hand of the model so much that you may as well just write by have, as you'll be quicker.

However, the code opus produces is good enough... So I rarely take action anymore in private projects

In which world throwing business logic to "optimize memory usage" isn't crappy solution? Or making 1T from 4G? It didn't returned decent solution, contrary it was worst than terrible

Guess who will be held responsible when your vibe-coded robot maims or kills someone.

I have enough experience writing code by hand and designing complex systems with a team working for me. In a sense, this is no different. When I had a half-dozen mid to senior level developers, I did not verify every line of code they wrote. I did not have any expectation that they would write perfect code, and they didn't.

But as the owner of the company, I was still liable for the product my employees made. That is the same thing here. It's not negligent to have Claude code write code for you at 10x the speed you can do it yourself. If I hadn't supervised my employees and put practices in place to control quality and safety, I would be liable. I'm actually far less worried about liability now, because I have my hands in every part of the system.

Calling it vibe coding is pejorative at this point, meant to imply that someone who has no skill or craft in software development just types a few prompts and gets something they didn't have a hand in the design or development. That's not what the professionals who are using AI coding are doing.

Can you share the prompt you're using for each model?

Sure thing! It is the same prompt for every model in the rollouts, here it is

  No tools are available. Do not imply that you searched, looked up, browsed, or verified anything externally. If the name is ambiguous, return distinct likely people or entities rather than blending them. Do not invent entries to fill the list. Return only JSON.

  Return fewer than 8 if fewer credible matches exist. Return {"results":[]} if you do not recognize any credible person or entity. Use this JSON shape:
  {
    "results": [
      {
        "rank": 1,
        "name": "Resolved person or entity name",
        "confidence": 0,
        "snippet": "Concise snippet supporting this result."
      }
    ]
  }

  Confidence is 0-100 for how strongly you recognize this specific person or entity. Snippet should be one short, complete search-result-style description (≤ 160 characters).

  The query is: Who is "<name>"?

The clusterer prompt is more intricate and I'm happy to share if of interest, but I have an invariant that every result showing up in a rollout must be clustered into one result (sometimes collapsed into the hallucinations section).

What if it doesn't return JSON?