Student here. Can someone give me one reason why I should continue in software engineering that isn't denial and hopium?
plumbing still looks like a safe choice for now.
Learning to Reason with LLMs
581–590 of 1001 posts
Re: Learning to Reason with LLMs
#582Student here. Can someone give me one reason why I should continue in software engineering that isn't denial and hopium?
What kind of student, at what kind of school?
Are your peers getting internships at FANGs or hedge funds? Stick with it. You can probably bank enough money to make it worth it before shtf.
Re: Learning to Reason with LLMs
#583Earlier quoted context omitted.
The calculator didn’t eliminate math majors. Excel and accounting software didn’t eliminate accountants and CPAs. These are all just tools. I spend very little of my overall time at work actually coding. It’s a nice treat when I get a day where that’s all I do. From my limited work with Copilot so far, the user still needs to know what they’re doing. I have 0 faith a product owner, without a coding background, can us…
Agreed. The sweet spot is people who have product owner skills _and_ can code. They are quickly developing superpowers. The overhead of writing tickets, communicating with the team and so on is huge. If one person can do it all, efficiency skyrockets. I guess it's always been true to some extent that single individuals are capable of amazing things. For example, the guy who's built https://www.photopea.com/ . But the…
Or people who can be product owners and can prompt LLMs to code (because I know him, that's me!).
I'm awestruck by how good Claude and Cursor are. I've been building a semi-heavy-duty tech product, and I'm amazed by how much progress I've made in a week, using a NextJS stack, without knowing a lick of React in the first place (I know the concepts, but not the JS/NextJS vocab). All the code has been delivered with proper separation of concerns, clean architecture and modularization. Any time I get an error, I can reason with it to find the issue together. And if Claude is stuck (or I'm past my 5x usage lol), I just pair programme with ChatGPT instead.
Meanwhile Google just continues to serve me outdated shit from preCovid.
Re: Learning to Reason with LLMs
#584Student here. Can someone give me one reason why I should continue in software engineering that isn't denial and hopium?
Three, though not slam dunks:
1. What other course of study are you confident would be better given an AI future? If there's a service sector job that you feel really called to, I guess you could shadow someone for a few days to see if you'd really like it?
2. Having spent a few years managing business dashboards for users, less than 25% ever routinely used the "user friendly" functionality we built to do semi-custom analysis. We needed 4 full time analytics engineers to spend at least half their time answering ad hoc questions that could have been self-served, despite an explicit goal of democratizing data. All that is to say; don't over estimate how quickly this will be taken up, even if it could technically do XYZ task (eventually, best-of-10) if prompted properly.
3. I don't know where you live, but I've spent most of my career 'competing' with developers in India who are paid 33-50% as much. They're literally teammates, it's not a hypothetical thing. And they've never stopped hiring in the US. I haven't been in the room for those decisions and don't want to open that can of worms here, but suffice to say it's not so simple as "cheaper per LoC wins"
Re: Learning to Reason with LLMs
#585This is incredible. In April I used the standard GPT-4 model via ChatGPT to help me reverse engineer the binary bluetooth protocol used by my kitchen fan to integrate it into Home Assistant. It was helpful in a rubber duck way, but could not determine the pattern used to transmit the remaining runtime of the fan in a certain mode. Initial prompt here [0] I pasted the same prompt into o1-preview and o1-mini and both c…
Very cool. It gets the conclusion right, but it did confuse itself briefly after interpreting `256 * last_byte + second_to_last_byte` as big-endian. It's neat that it corrected the confusion, but a little unsatisfying that it doesn't explicitly identify the mistake the way a human would.
Re: Learning to Reason with LLMs
#586Earlier quoted context omitted.
I guess they'd also charge for the chain of thought tokens, of which there may be many, even if users can't see them.
That would be very bad product design. My understanding is that the model itself is similar to GPT4o in architecture but trained and used differently. So the 5x relative increase in output token cost likely already accounts for hidden tokens and additional compute.
> While reasoning tokens are not visible via the API, they still occupy space in the model's context window and are billed as output tokens.
https://platform.openai.com/docs/guides/reasoning
So yeah, it is in fact very bad product design. I hope Llama catches up in a couple of months.
Re: Learning to Reason with LLMs
#587I m wondering, what kind of "AI wrappers" will emerge from this model.
Re: Learning to Reason with LLMs
#588Student here. Can someone give me one reason why I should continue in software engineering that isn't denial and hopium?
There's so much software yet to to be written, so much to automate, so many niches to attack that you need not worry. It takes humans to know where to apply the technology based on their heart, not brains. Use AI in the direction only you can ascertain; and do it for the good of HUMANITY. It's a tool that makes the knowledge posterity has left us accessible, like mathematics. Go forth an conquer life's ills young man…
Machines don't really "know" anything they just manipulate what is already known; Like a interactive book. It's just that this AI book is vast.
Re: Learning to Reason with LLMs
#589First shot, I gave it a medium-difficulty math problem, something I actually wanted the answer to (derive the KL divergence between two Laplace distributions). It thought for a long time, and still got it wrong, producing a plausible but wrong answer. After some prodding, it revised itself and then got it wrong again. I still feel that I can't rely on these systems.
Look where you were 3 years ago, and where you are now. And then imagine where you will be in 5 more years. If it can almost get a complex problem right now, I'm dead sure it will get it correct within 5 years
I'm not? The history of AI development is littered with examples of false starts, hidden traps, and promising breakthroughs that eventually expose deeper and more difficult problems [1].
I wouldn't be shocked if it could eventually get it right, but dead sure?
Re: Learning to Reason with LLMs
#590Just added o1 to https://double.bot if anyone would like to try it for coding.
---
Some thoughts:
* The performance is really good. I have a private set of questions I note down whenever gpt-4o/sonnet fails. o1 solved everything so far.
* It really is quite slow
* It's interesting that the chain of thought is hidden. This is I think the first time where OpenAI can improve their models without it being immediately distilled by open models. It'll be interesting to see how quickly the oss field can catch up technique-wise as there's already been a lot of inference time compute papers recently [1,2]
* Notably it's not clear whether o1-preview as it's available now is doing tree search or just single shoting a cot that is distilled from better/more detailed trajectories in the training distribution.