kaspermarstal

I switched from Opus 4.6 -> Opus 4.7 -> GPT 5.5 and tried Flash 3.5 tonight and I was not impressed. It is straight up unreliable, e.g. deleting code and forgetting to add the new stuff it was asked to, then happily marking the task as complete with up-beat conclusion. I personally appreciate GPT 5.5 toned-down, objective style so really dislike how this model feels. I get that it's a flash model and not in the same league as GPT 5.5 but their marketing suggest otherwise so thy are just setting themselves up for disappointment.

I’m quite sure that’s the en state of software except without the software around it. There will only be an AI and interface. For now, though, while tokens cost a non-trivial amount of energy, I think you can do something more useful if you have the LLM modify the program at runtime because it’s just may orders of magnitude cheaper. Fx, use the BEAM, it’s actor model, hot code reloading, and REPL introspection and you can build a program that an LLMs can change, e.g. user says “become a calculator” and “become a pdf to html converter”.

I’m not just making this stuff up of course, got the idea yesterday after reading Karpathy’s tweet about Nanoclaws contribution model (don’t submit PRa with features, submit PRs that tell an llm how to modify the program). Now I can’t concentrate on my day job. Can’t stop thinking about my little elixir beam project.

Cool, thanks for sharing!

Can you share technical details please? How is this implemented? Is it pure prompt-based, plugins, or do you have like script that repeatedly calls the agents? Where does the kanban live?

Ooooh, graphs that goes up! I want that.

Looks really cool, great job.

The Wolf maintainer has done heroic work ...

I commend the fact they acknowledge the maintainer's work, but seeing the singular 'maintainer', I can't help but notice the weight on that one person's shoulders.

Keyboard Princess is good, Artificial Manager is even better.

Techno bureaucrating is unironically a better term than vibe coding.

Naming things is hard for a reason. It shapes how we think and communicate. When I'm pointing an LLM at PostgreSQL internals vs. asking it to make a navbar prettier, I'm collaborating in fundamentally different ways. Having distinct terms helps us share ideas and set appropriate expectations.

My point exactly, it is not vibe coding so it should not be called vibe coding. What should we call it then?

Cool, I did not know that. That makes perfect sense.

We need a new term for LLMs actually solving a hard problems. When I help Claude Code solve a nasty bug it doesn’t feel like “vibing” as in “I tell the model what I want the website to look like”. It feels like sniping as in “I spot for Claude Code, telling how to adjust for wind, range, and elevation so it can hit my far away target”.

Thanks, appreciate it. Indeed, and Anthropic did something similar for Google sheets a year ago. I am dying to know why they decided this should not be part of their excel effort. They obviously put a lot of work and thought into claude for excel so it must be intentional.

Anyone from Anthropic here that would like elaborate?

If someone is processing 50k rows, that means they found real value and the UX is working. That's the whole point.

Also, 50k rows wouldn't cost $50k. More like $100 with Sonnet 4.5 pricing and typical numbers of input/output tokens. Imagine the time needed to go through 50k rows manually and math doesn't really work for a horror story.

So cool, I hope they pull it off. So many people use Excel. Although, I always thought the power of AI in Excel would come from the ability to use AI _as_ a formula. For example, =PROMPT("Classify user feedback as positive, neutral or negative", A1). This would enable normal people (non-programmers) to fire off thousands of prompts at once and automate workflows like programmers do (disclaimer: I am the author of Cellm that does exactly this). Combined with Excel's built-in functions for deterministic work, Claude could really kill the whole copy-pasting data in and out of chat windows for bulk-processing data.

Models are less verbose, so produces fewer output tokens, so answers cost less.

Okay, I'll bite. There are many clever people out there who figured out that spreadsheets is a great UI paradigm for building low-code AI workflows. That is good for users. And I find it is good for me. Thanks for validating the idea guys.

I chuckled

Hey, sure! Send me an email at kasper at getcellm dot com and let’s find a time

Looks very similar yes, it is a great UI paradigm for running many prompts. I think of spreadsheets as the OG low-code tool and with just a sprinkle of LLMs, people can do so much more with tools they already know

Yes it does, you can toggle it on and off. Send me an email at kasper at getcellm dot com or sign up to the waitlist on getcellm dot com and I will personally onboard you!

Cool, thanks!

No not yet, we are about to onboard the first users on the waitlist one-by-one and when we have ironed out the major issues that we will inevitably discover, we will open up for paid users after that.

I wouldn’t worry too much about missing out, as you probably very well aware, whatever you choose to work on takes incredible amounts of time and energy to get off the ground. Now you just have more time to out into something else :)

Yes exactly like that!

I'm continuing the work on Cellm, an Excel extension that let's you call LLMs in cell formulas like =PROMPT(A1, "Rate the sentiment of the customer feedback as positive, neutral, or negative"), and then drag the formula down to apply the same prompt to thousands of rows. I built it after my girlfriend had to manually classify 7,500 research papers. Cellm automates that kind of repetitive work.

Since we added MCP and the use of structured output to "spill" multiple return values into adjecent cells, it is the quickest way I know of to monitor competitors blogs everyday before my 09:00 meeting. And also the quickest way I know of to test new AI models. I have a sheet with SimpleQA, MMLUPro, or GPQA Diamond and testing a new model is a matter of adding a new column. The whole idea is to enable normal people (like, non-techies) to automate manual, repetitive tasks with AI like programmers routinely do.

https://github.com/getcellm/cellm

Interesting. Can you pitch me why I should migrate from cloudflare R2 to you?

Made me click it though

Glad you tried it out! Excel triggers recalculation when a referenced cell updates, just like with any other formula. This is also why responses are not streamed, as every update would trigger recalculation. But if the async behavior of responses messes with the recalculation logic I am very interested in looking into it and you are most welcome to open an issue in the repo with steps to reproduce.

Yes, and it will be

I am not going to claim or report any kind of accuracy, especially with such a small model and such a specific, context dependent use case. It is the user’s responsibility to cross validate if it’s accurate enough for their use case and upgrade model or use another approach if not.