They tested it in February 2024. There are presumably privacy reasons limiting them to open models. It probably was the most promising at the time for their use case.
It was released roughly a year ago, I wouldn't really say this is an issue. While there has been plenty of progress in the meanwhile, actually conducting research takes time and effort, you can't expect everything to be done on the very bleeding edge at all times.
5 people evaluating 45 responses? This doesn't take a year to do. The issue is that this study is was poorly funded and slow - model development has far outpaced the results and there's likely little to no longevity in the outcome.
According to the linked PDF which includes a report from AWS professional services, the PoC was originally devised in September 2023 and conducted in January/February of 2024, before the release of Llama 3. They tested Llama2-70B, Mistral-7b and MistralLite. They didn't evaluate proprietary models such as GPT4 probably because they would've wanted to be able to deploy it within Australia or with an Australian company
I find it weird anyone is surprised by this. LLMs don't understand what they're doing at a fundamental level, so when you ask for a summary, you're asking for it to make something more concise. Which it will proceed to do, with no knowledge of the relative importance of what it's pruning out.
There's a whole category of issues around this that I don't see how the current formulation of AI based on LLMs can solve.
I find it weird anyone believes the Earth is flat, yet here we are. Most people don’t understand or care how LLMs work and the companies pushing those systems have little interest in explaining. The state we’re in is sad and disappointing, but far from surprising.
> LLMs don't understand what they're doing at a fundamental level
There are a large number of people on this website, which is ostensibly a geek forum, who fail to understand even that simple concept. Worse, they actively refuse the notion. What chance do non-technical people have?
Summarization is one of those key functionalities of LLMs that laypeople can also easily understand and relate to.
I think this article also underpins the hunch that results in ROUGE and similar NLP benchmarks are not necessarily a guarantee for good performance in sector-specific summarization tasks, where the expectations are defined by human users for a specific domain.
It reminds me of the Stanford study on the use of LLMs for legal QA including the eye-wateringly expensive legal-specific LLMs:
https://hai.stanford.edu/news/ai-trial-legal-models-hallucin...
This feels a _bit_ like 'water is wet, study finds'. Like, I'm a little surprised that they felt the need to run the trial; even the most overconfident AI booster would probably have been reluctant to tell them they'd find anything different.
What would actually help is a plot that shows how the versions of Llama models are getting better (or not) at this summarization task, and then attempt to estimate when they will reach a human standard. This would allow us to see things in perspective.
16 comments
[ 0.20 ms ] story [ 59.4 ms ] threadThis is an old model that was not dominant even when released. This study must be fairly old or I question the qualification of the group running it.
5 people evaluating 45 responses? This doesn't take a year to do. The issue is that this study is was poorly funded and slow - model development has far outpaced the results and there's likely little to no longevity in the outcome.
There's a whole category of issues around this that I don't see how the current formulation of AI based on LLMs can solve.
I find it weird anyone believes the Earth is flat, yet here we are. Most people don’t understand or care how LLMs work and the companies pushing those systems have little interest in explaining. The state we’re in is sad and disappointing, but far from surprising.
> LLMs don't understand what they're doing at a fundamental level
There are a large number of people on this website, which is ostensibly a geek forum, who fail to understand even that simple concept. Worse, they actively refuse the notion. What chance do non-technical people have?
So not cheaper, faster, or more consistent?
Sounds more like "worse than humans in the few ways measured in this limited trial"
Better than an average human? Absolutely.