maybe ask ChatGPT....
The AI is taking our jerbs thread
- Thread starter mid_gen
- Start date
I've had numerous instances of various models recommending deprecated code, or just making it up - this has improved though.
There was one case where they all kept coming up with different versions of the wrong answer - I eventually found the 7 year old repo that they had been trained on.
I’ve already asked if it wants to ‘kill all humans’.
not really. LLM simply provide a probability distribution over tokens given a sequence of tokens, a token is then selected stochastically based on the distribution and temperature params. Then the whole sequence is fed back in to generate 1 more token.
So ot can provide slightly different text to the same input due to the random sampling. But it should also be consistent. E.g if you ask what is the capital of France it should always be confident and correct in Paris due to the probability distribution, but it could rephrase the answer slightly differently due to randomness.
When you get wildly different outputs it is often due to hallucinations.
So... what I said was true then, it gives different answers by design.
To ask it a general question like "where is nice to go on my holidays?" works differently to "what is the capital of France?" The latter is where the hallucinations (mostly) occur or are at least (mostly) noticed as there is a factual answer than can be right or wrong. It telling you to go to Norway instead of Morocco isn't as much of an issue. But this is all sort of a distraction when yes, they hallucinate way too much and stopping it is a big headache.
To ask it a general question like "where is nice to go on my holidays?" works differently to "what is the capital of France?" The latter is where the hallucinations (mostly) occur or are at least (mostly) noticed as there is a factual answer than can be right or wrong. It telling you to go to Norway instead of Morocco isn't as much of an issue. But this is all sort of a distraction when yes, they hallucinate way too much and stopping it is a big headache.
Last edited:
Soldato
- Joined
- 30 Sep 2005
- Posts
- 16,788
There was a thread on here many years ago where the general concensus at the time was that there would be no loss of jobs. People would retrain and move sectors/skills.
Will be interesting to see how/if that view changes over time.
Will be interesting to see how/if that view changes over time.
Plato conjecture is that when all models get to a certain size they essentially become the same model.
So that has ramifications on businesses that use AI, where they will struggle to differentiate to create value if they only use the standard large models.
That's been happening for years. The whole M&S debacle is also a good lesson in getting what you pay for.
That's what I mean.
All this talk about "AI taking our jobs" we really haven't see that in action. But we still see companies moving teams to cheaper countries and not moving them towards AI to reduced costs.
They're definitely trying - corporations are all about maximising profits even if it's at the expense of quality, so it will be the same thing as with outsourcing; some companies will push for it and maybe they're lucky enough to avoid issues, but over the years there'll definitely be a degradation in quality.
Companies are laying off people, and sometimes they will mention AI. But when they mention AI it is often just that they are investing in different areas including AI, so AI is is merely causing a change in where the headcount is. When companies have more directly claimed efficiency/automation from AI has allowed them to reduce recourses there is never any evidence it is anything other than a standard layoff to reduce costs. Companies are very keen to boast about how their investments in AI are coming to fruition, or to simply provide a message that They aren't simply doing layoffs to increase C-suite bonuses but magical AI is making their business so much better they can lay off thousands. No way is Co-pilot actually replacing the 5000 engineers Microsoft laid off
On that front it is all relative and completely unrelated to the country.
The best developers in India for example are equally good if not better than the best developers at the same grade in the US or UK. The mediocre engineers are the same everywhere.
The only difference is the relative salaries. You could pay 250k USD for a top engineer in the US or Switzerland, but that gets you 2-3 of the same caliber in UK and 5-6 in India.
There US/UK companies can easily reduce software dev costs and even increase the quality by hiring overseas.
But if companies pay bottom dollar overseas they get the same quality as bottom barrel engineers in the UK. If you want a good software engineer in India it costs about £50k GBP, and you will get someone who walks circles around a UK engineer on 6 figures. Conversely you can replace you £35k very junior engineer in the UK with some on 10k on India and expect the same minimal output.
Yea but the outsource places generally don't have that level of developer. It's usually a buy 1 get 1 free situation, and you're getting someone who's been put through a handful of courses to have just enough base knowledge.
I'm sure there are highly qualified developers that produce quality software, but they're not working at these mass outsourcing companies.
Last edited:
Like everything, the outsourcing companies have multiple levels and different companies specialize at different price points.
The problem is not the outsourcing, the problem is simply companies trying to save by paying peanuts. The companies that hire the lowest grade developers in the UK will have terrible software quality, these same penny pinchers will look to outsource to India, but instead of paying the same salaries as UK software engineers but getting far better developers, it is done entirely to reduce costs. So they can get the same poor developers at a lower cost as the peanut eating monkeys back in Britain . Blaming outsourcing is just a scapegoat for insufficient investment.
It is a lot harder to convince some that they can pay the same salaries for indian developers as UK but increase the quality and productivity significantly and see future ROI by having higher quality software. Software quality is hard to measure so is often ignored by executive leadership. You end up with all kinds of useless metrics like defect density
The funniest I have seen id cyclcometric complexity which i saw 1 leader claim needs to be as high as possible because it is a sign of experienced engineers making more sophisticated code that junior devs can't achieve....
Essentially, companies that just want the lowest cost produce terrible software where ever the engineers are located.
Soldato
- Joined
- 21 Apr 2007
- Posts
- 6,627
Going to be frank, every time I see this I just think it's full on cope.
I've also seen it compared it to the industrial revolution where AI somehow makes more jobs, how exactly? as an AI researcher or one of handful of elite devs that work for open AI?
Sure.
The AI Con by Emily M Bender and Alex Hanna review – debunking myths of the AI revolution
Will new technology help to make the world a better place, or is AI just another tech bubble that will benefit the few?www.theguardian.com
AI is bad for workers good for tech companies and plagiarists. Basically zero sum gain for everyone else.
Indeed.
We're being told that it's AI first. Not augment.
The human race is not AI first.. it's Humans First.
There's an academic paper that basically says AI is following Plato's Representation Hypthesis: https://arxiv.org/pdf/2405.07987
This means that all models, as they get big become the same.
This has massive implications as we end up with no differentiator and no real innovation from the same predictions and output of AI.
My view as an extension, therefore each AI needs to have a coded difference of opinion.. and that's where AI wars start as the AIs will then need to be able to compromise.
The reality is that the universe as a large set is the set and there's no way to reinterpret it without being hypothetical.
This means that all models, as they get big become the same.
This has massive implications as we end up with no differentiator and no real innovation from the same predictions and output of AI.
My view as an extension, therefore each AI needs to have a coded difference of opinion.. and that's where AI wars start as the AIs will then need to be able to compromise.
The reality is that the universe as a large set is the set and there's no way to reinterpret it without being hypothetical.