31–40 of 62 posts
Not for RAG and Automation applications But for Fixing Hallucinated code that written by Non Software Engineers.
That is completely contingent on if present the day SWEs stop working on trying to eliminate the hallucinations.
Eh, it's only a matter of time until new nvidia clusters are in production and we start seeing what hardware designed for training massive models (roughly 100x bigger than anything we know of at least) can do. With some additional architectural changes we might start seeing some sparks of real human stupidity.
The exponential graph people are posting is just to pump the stock price.
This article feels like stating an opinion. I suppose the autopilot analogy is at the heart of their argument but the argument as of now feels a bit too flat. There are definitely some jobs that technology completely displaced, and there are some that merely made something more accessible. For example, grinding grain with the advent of modern milling equipment vs microwaves. Milling machines make hand driven milling…
> Milling machines make hand driven milling something of a niche whereas microwaves arguable have displaced no jobs in cooking and food preparation. I would argue that generative AI in the hands of an already skilled developer is equivalent to the mill scenario. One person's expertise leveraged by way of this tool can allow for output that is orders of magnitude greater. The mere inertia of the "machine" encourages m…
AI isn't about to become capable of doing the job of a software developer any time soon, not least because two of it's major weaknesses - extremely limited reasoning ability and extremely limited learning ability (in-context only) are the two things that software developers most heavily lean on. It's not clear whether these are things that can really be addressed without major architectural change. There's still mass…
If you're thinking over the next 3-5 years, then AI is a novelty in most cases, a useful tool in some cases, and something truly disruptive in a fairly narrow set of cases.
If you expand the time frame to 10-20 years, then AI is poised to become something that truly upsets most of the economy. Obviously, nothing can be said for certain, there could be black swan events, etc. But on the current trajectory of cheaper, more powerful compute and better understanding of the engineering around getting useful results out of AI, you'd be foolish to think any white collar profession won't be deeply affected by the technological shift that's coming.
Earlier quoted context omitted.
That is completely contingent on if present the day SWEs stop working on trying to eliminate the hallucinations.
I'm curious as to what engineering solutions you think are available to solve this problem? All the solutions I've heard up until now is "we just need more data" which isn't particularly sustainable or something a SWE would be in charge of doing
AI is exponentially improving and there is no reason to believe that this won’t continue.
Yes, but it will strongly decrease demand for new ideas. AI is a "win more" technology for current software architecture incumbents. The more popular a tech is, the more content there is about it, the more capable an LLM is of giving a usable answer. A programmer who relies on LLMs for coding assistance will be far less likely to try something new, and it will have a chilling effect on the adoption of all new tools,…
I for one used to hire people outside Germany for smaller projects, doing data analysis, writing interfaces, migrating data. With all the new tools I don't do this anymore. I can come up with solutions that are good enough, and I don't have the overhead with back and forth emailing, writing requirements, getting a result, giving feedback, etc. This is me, n=1.
If not, your competition will.
AI isn't about to become capable of doing the job of a software developer any time soon, not least because two of it's major weaknesses - extremely limited reasoning ability and extremely limited learning ability (in-context only) are the two things that software developers most heavily lean on. It's not clear whether these are things that can really be addressed without major architectural change. There's still mass…
What kind of time frame is "any time soon"? If you're thinking over the next 3-5 years, then AI is a novelty in most cases, a useful tool in some cases, and something truly disruptive in a fairly narrow set of cases. If you expand the time frame to 10-20 years, then AI is poised to become something that truly upsets most of the economy. Obviously, nothing can be said for certain, there could be black swan events, etc…
This feels like a lot of "20 years away" type predictions. There needs to be some sort of step change in the technology, and perhaps more than one or on more than one front to make this possible but its far enough away that 20 years feels like a safe bet.
When people say this about LLM based AIs, I wonder how they imagine text prediction making those step changes. Is your thinking that text prediction is generalization enough that doing more of it better, faster, and on scaled hardware is sufficient to hit a tipping point?
Earlier quoted context omitted.
I'm curious as to what engineering solutions you think are available to solve this problem? All the solutions I've heard up until now is "we just need more data" which isn't particularly sustainable or something a SWE would be in charge of doing
It’s not my expertise. However, if it’s a barrier to adoption it’s a problem to solve and will be. AI is exponentially improving and there is no reason to believe that this won’t continue.
Not for the last two years at least. Over the last year it's been obvious that there are diminishing returns already...
We increase the computational power by a factor of ten, and see a factor of 1.5 improvement.