6 ms·
If you allow LLMs (Large Language Models) to directly call out to APIs, like POSIX, Vulkan, Win32, X11, etc. then it may become possible to program things using
by ogogmad 4y ago
If you allow LLMs (Large Language Models) to directly call out to APIs, like POSIX, Vulkan, Win32, X11, etc. then it may become possible to program things using a "Socratic" method that involves producing no code at all. You simply channel inputs to the LLM, collect API calls from the LLM, and execute the API calls it makes. This will create a feedback loop that'll eliminate the need to produce code at all. Code can then be produced selectively as an optimisation, like what we call "compilation" today.
Of course, it's currently unreliable and slow, but the trend is a fast one.
Additionally, it might be given a memory store it could read from and write to -- and a StackOverflow-like network to curate advice from other AIs like itself.
A half-way version of this might soon start to exist.
- ilyt 4y agoAnd we will be able to blame production melting down on "AI" misinterpreting what someone said. "I told it to delete users from my contact list, not just delete them from system!"
- dragonwriter 4y agoFortunately, the business user to analyst to dev game of telephone with people never results in misinterpretations that survive to production.
- ehutch79 4y ago100% someone would say 'delete the users from my contact list in the system'
- mjburgess 4y agoThe problem with these generative systems is that they're synthesizing existing code and their output isnt therefore always valid. Ie., they arent programming (which involves following discrete constraints of the language). They are, therefore, largely tools for power-users. You need to understand why codepiolt/etc. is suggesting a completion before its actually useful. If you prefer: Intelligence is knowing what question to ask, and knowing how to validate the answer.
- anotheracctfo 4y agoOh god no. No please don't let "power-users" anywhere near a Low Code framework. It doesn't take long for them to fuck up and go ask a real developer what is wrong. I've seen things...
- dragonwriter 4y ago> It doesn't take long for them to fuck up and go ask a real developer what is wrong. I've seen things... Worse is when it doesn’t take them long to fuck up, but it does take them (or, often, other people in the org, possibly after they’ve left) long to bring in a dev... I’ve seen things, too...
- tartoran 4y agoThe sooner the better. Would you prefer stage one cancer or stage four cancer?
- dragonwriter 4y ago> The problem with these generative systems is that they're synthesizing existing code and their output isnt therefore always valid Actual programmers have this problem too, though they often use non-AI tooling to catch and prompt them to correct invalid code. No reason a system using an LLM for code generation couldn't do that, too. Anticipating the “but even if it is valid it won’t always be correct” followup: well, valid human-written code also often has bugs.
- wongarsu 4y agoAt that point, the LLM is just a compiler, and your prompt becomes the "code". I can easily imagine something like this taking off. After all it would be a giant step towards two things we have been trying since the 60s: a programming language intuitive enough that business people can use it, and a declarative approach where you describe what you want to achieve, not which steps to take. That said, if all the current prompt engineering is anything to go by, using such an approach effectively might still require a lot of skill. But the barrier to entry would be a lot lower, and for some classes of problems it might be much more efficient in terms of programmer time.
- thesuperbigfrog 4y ago>> You simply channel inputs to the LLM, collect API calls from the LLM, and execute the API calls it makes. What could possibly go wrong? I know! Hook it up to manage people's retirement savings / 401(k) for pennies on the dollar compared to traditional portfolio managers.