I’m well aware, thank you, have been coding for 40+ years (including 6502 and 68000 assembly), masters in computer science, have built healthcare software where bugs can lead to death. But with LLMs enabling us to generate source code faster, our review process is becoming an increasingly larger bottleneck for productivity. We need to start thinking how we can scale this process.
So if you don’t have a long term memory, you’re not capable of sentience? Like the movie memento, where the main character needs to write down everything to remind him later because he’s not able to remember anything. This is pretty much like llms using markdown documents to remember things.
For practical every day uses, does it really matter if it is "real thinking" or just really good "artificial thinking" with the same results? The machine can use artificial thinking to reach desired goals and outcomes, so for me it's the kind of thinking i would want from a machine.
Most AI tooling is shipped with a feedback loop around the LLM. The quality of Claude Code for example lies in the feedback loop it provides on your code. Maybe the LLM itself isn't thinking, but the Agent which ships an LLM plus feedback loop definitely shows thinking qualities.
Just now in an debugging session with claude code:
* let me read this file...
* let me read this file...
* I think there's a caching issue with the model after dropping the module. Let me check if there's a save or reload needed after DROP MODULE. First, let me verify something:
* creates a bash/javascript script to verify its assumption
* runs the script (after review and approval)
* Aha! I found the problem! Look at the output...
An LLM by itself is not thinking, just remembering and autocompleting. But if you add a feedback loop where it can use tools, investigate external files or processes, and then autocomplete on the results, you get to see something that is (close to) thinking. I've seen claude code debug things by adding print statements in the source and reasoning on the output, and then determining next steps. This feedback loop is what sets AI tools apart, they can all use the same LLM, but the quality of the feedback loop makes the difference.
Instead of downvotes i would appreciate some insightful comments on this, as i'm currently struggling with this problem. In the last week i've vibe-code (vibe-engineered?) a typescript project with 230+ commits, 64 typescripts files, with 27k+ lines of code. Too much to actually read. Validation mostly through testing, automated test, architecture reviews (generate mermaid diagrams). I'm mostly reviewing the code structure and architecture, libraries it uses, etc. It has 600+ unit and integration tests, but even reviewing those is too much...
AI code generators are getting better fast, in the near future they will be able to produce good changes faster than you can review. How will you deal with it then? Most vibe coding tools can also produce smaller PR, but then you have to deal with 250+ PRs in 1 week. Is that more manageable? My guess is we need new tool, get the human out of the loop. More automated reviews, tests, etc.
Going forward it will probably mostly be a combination of both. Similar to how CAD software is also using ML and AI models for simulations and other tools.
Skills are just lazy loaded prompts. The system prompt will include the summary of every skill, and if the LLM decides it want to use a specific skill the details of that skill will be added to the system prompt. It's just markdown documents all the way down.
The article seems a few months too late. Claude (and others) are already doing this: i've been instructing claude code to generate code following certain best practices provided through URLs or asking it to compare certain approaches from different URLs. Claude Skill uses file "URLs" to provide progressive disclosure: only include detailed texts into the context if needed. This helps reduce context size, and improves cachability.
We're now getting in a position where we have CAD for software, aka CASE, Computer Aided Software Engineering. You can focus on the design of the software, instead of spending hours typing out code.
Oracle has a nice way to bundle stores procedures in packages, which makes large amounts of stored procedures manageable. So still ahead of Postgres, but Postgres is definitely good enough.
Bosses should be able to make decisions about jobs or AI. That ok.
But as a society we have to ask ourselves if replacing all jobs with AI will make for a better society.
Life is not all about making as much money as possible. For a working society citizen need meaning in their lives, and safety, and food, and health. If most people get too little of this, it may disrupt society, and cause wars and riots.
This is where government needs to step in, uncontrolled enterprise greed will destroy countries. But companies don´t care, they'll just move to another country. And the ultra-rich don´t care, they'll just put larger walls around their houses or move country.
Just had a chat with AI to see how we could address the issues mentioned in the article. You can create models that cater to multiple use cases. You can split the domain model into facts (tables) and perspectives (views). This gives you a lot of flexibility in addressing the different perspectives presented in the artcile from a shared domain model.
But with LLMs is there really more to understand? They’re just large functions that take numerical input and transform it into numerical output based on trained weights. There is nothing behind the scenes doing things we don’t understand. The magic is in the weights, and we know how to create these based on training data.
Regarding the car, if you know how to build a car, you understand how a car works. A driver is more like someone using and llm, not a developer able to create an llm.
Yes, I rather bring an iPad than a laptop, easier lighter and smaller to travel with, also better for movie watching in the plane. Laptop is hard to fit on a table.
Impressive, it clearly is able to read the score, see patterns, timing, chords, and apply music theory to it. Would be interesting to give it editing and playback capabilities, e.g., by connecting it to something like strudel.
This article describes how Belgian supermarkets are replacing music played in stores by AI music to save costs, but you can easily imagine that the ai could also generate music to play to the emotions of customers to maybe influence their buying behavior: https://www.nu.nl/economie/6372535/veel-belgische-supermarkt...
Did it show you the reasoning? Did it recognize the notes, the scale, tempo and determine the emotion effect of these or did it use some other reasoning?
We don’t know how to build humans from scratch (other than letting nature do it) so that’s not really a relevant example. We only know how to fix certain aspects of defect humans, and how to minimize the chance of future defects, but that’s far from building a human from scratch.
So just because all the parts are collaborating to get the desired outcome, and specific aspects of that outcome cannot be contributed to specific parts of the llm you think we don’t understand LLMs? With mixture of expert systems we’re introducing dedicated subsystems into the llm responsible for specific aspect of the llm, so we’re partially moving in that direction.
But overall in my opinion if devs are able to rebuild it from scratch with a predefined outcome, and even know how to improve the system to improve certain aspects of it, we do understand how it works.
Not true, it’s a pretty good machine to edit videos on, extra cpu/gpu power is welcome there. I connect it to an external monitor, keyboard and mouse and use Final Cut Pro to edit videos. On holidays I can connect my action cam to the iPad, download the videos and edit them, without having to bring a laptop.
There's no magic involved, the LLM creators can go anywhere and rebuild an LLM with pretty much the same outcome, if they have the same training data. With unlimited time you could even reproduce the output of an LLM manually as it is just a lot of mathematics. Including reasoning, as that is mostly adding words in the context that will steer the word predication to include reasoning. As this is a useful LLM behavior this context is now part of the training data, so it becomes part of the neural network weights.
I think there are people who know exactly how it works, they know how neural networks work, they know how the transformer architecture works, how attention works, embeddings, tokenization, etc. We just can’t define the weights of the connections between the neurons.
This being text based makes it really easy to have AI generate the music. Now waiting for Strudel agent that will transcribe music into strudel notation.