HN user
My problem with this kind of work is—-obviously they do. Did anyone seriously think otherwise? I’m shocked why these are even questions deserving scientific scrutiny. Have people truly lost their critical thinking that badly already?
Thanks!
Sorry, but these people are not victims. I went through a tech PhD; it was well-known how fast the wind changes and the trendy topic falls by the wayside. Big data and crowdsourcing pre-AI boom, then ML, then ethics of AI (around 2020-21), now we are at AI agents which is inherently more product-oriented. Beyond this, we had COVID and a now brutal tech market, despite the US economy seemingly way up and the AI craze continuing. If you went into a CS PhD thinking nothing changes and people will beg you to take cushy salary positions with total freedom, then you simply didn’t do your research.
Wow. This actually disproves a key subtext of the match mentioned by some commentators: that Ding failed to convert winning positions to wins. Instead, it shows that Ding converted more often than Gukesh. The fact that Gukesh won seems more a statistical anomaly in light of this evidence. We are indeed probably post-hoc rationalizing the winner.
“Do politics have artifacts?” was the rejoinder article. IMO that article should be as widely read as the main one, because it provides a warning to those who take the main one as gospel. Link: https://journals.sagepub.com/doi/abs/10.1177/030631299029003...
This seems like a sheets implementation of something like ChainForge (https://github.com/ianarawjo/ChainForge).
It's curious that Anthropic is entering the LLMOps tooling space ---this definitely comes as a surprise to me, as both OpenAI and HuggingFace seem to avoid building prompt engineering tooling themselves. Is this a business strategy of Anthropic's? An experiment? Regardless, it's cool to see a company like them throw their hat into the LLMOps space beyond being a model provider. Interested to see what comes next.
The original poster making this claim used a t-test to compare means (https://x.com/RobLynch99/status/1734278713762549970?s=20). Turns out the data is not normally distributed, making a t-test worthless (https://www.statology.org/t-test-assumptions/).
There might be other tests to do, but for this specific setup, the claim has been debunked.
Can’t reproduce this. See for yourself: https://x.com/IanArawjo/status/1734307886124474680?s=20
Inspectable evaluation flow in ChainForge: https://chainforge.ai/play/?f=2yvqkpe1vpus8
This seems like a sheets implementation of something like ChainForge (https://github.com/ianarawjo/ChainForge).
It's curious that Anthropic is entering the LLMOps tooling space ---this definitely comes as a surprise to me, as both OpenAI and HuggingFace seem to avoid building prompt engineering tooling themselves. Is this a business strategy of Anthropic's? An experiment? Regardless, it's very cool to see a company like them throw their hat into the LLMOps space beyond being a model provider. Interested to see what comes next.
ChainForge lets you do this, and also setup ad-hoc evaluations with code, LLM scorers, etc. It also shows model responses side-by-side for the same prompt: https://github.com/ianarawjo/ChainForge
Thanks!
There is a long term vision of supporting fine-tuning through an existing evaluation flow. We originally created this because we were worried about how to evaluate ‘what changed’ between a fine-tuned LLM and its base model. I wonder if Vertex AI has an API that we could plug-in, though, or if it’s limited to the UI.
Hey Eric! Thank you! As an aside, we are looking to interview some people who’ve used ChainForge (you see, we are academics who must justify our creations through publications… crazy, I know). Would you or anyone on your team be interested in a brief chat?
Can contact me here: https://twitter.com/IanArawjo Or find email on CV here: ianarawjo.com
At any rate, glad it was helpful!
Thank you for the kind words! Looking at the photo, I think you wouldn’t need the last prompt node there.
As far as evaluating functions go, that’s unfortunately a ways off. But, we generally prioritize things based on how many people posted GitHub Issues about it/want it. (For instance, Chat Turn nodes came from an Issue.) If you post a feature request there, it’ll move up our priority list, and we can also clarify what the feature precisely should be.
We don’t view it as a competition here. These other tools are for LLM app building, which is great, but isn’t the focus of ChainForge. Instead, we’re focused on helping devs find the right prompt, and evaluate and inspect LLM outputs. So while we might visually look similar, the goals are rather different.
Some CF users, for instance, might not be app builders at all —they just want to audit models.
I think both problems —prompt engineering and LLM app building —are hard, and deserve their own dedicated tools.
ChainForge looks similar visually, but is very different in practice. We target evaluating and inspecting LLM outputs, rather than building LLM applications. So, some things are certainly easier to do in CF, while others will certainly be easier in other tools that target LLM app building.
This is a pretty cool idea. One comment —the print representation should be entirely akin to a dict, just braces. No DotDict wrapper. That’d be cool.
Thanks for the clarification! Yes, I see now that auto-evals here is more AI agent-ish, than a one-shot approach. Still has the trust issue.
For suggestions, one thing I'm curious about is how we can have out-of-the-box benchmark datasets and do this responsibly. ChainForge supports most OpenAI evals, but from adding this we realized the quality of OpenAI Evals is really _sketchy_... duplicate data, questionable metrics, etc. OpenAI has shown that trusting the community to make benchmarks is perhaps not a good idea; we should instead make it easier for scientists/engineers to upload their benchmarks and make it easier for others to run them. That's one thought, anyway.
What does 'top 50%' responses mean here, though? You'd need to have a ground truth of how 'good' each score was to calculate that --and if you had ground truth, no need to use an LLM evaluator to begin with.
If you mean trusting the LLM scores to pick the 50% 'top' responses they grade, this doesn't get around the issue of overly trusting the LLM's scores.
I like the support for Vector DBs and LLaMa-2. I'm curious as to whether and what influences compelled PromptTools, and how it differs from other tools in this space. For context, we've also released a prompt engineering IDE, ChainForge, which is open-source and has many of the features here, such as querying multiple models at once, prompt templating, evaluating responses with Python/JS code and LLM scorers, plotting responses, etc (https://github.com/ianarawjo/ChainForge and a playground at http://chainforge.ai).
One big problem we're seeing in this space is over-trust in LLM scorers as 'evaluators'. I've personally seen that minor tweaks to a scoring prompt can sometimes result in vastly different evaluation 'results.' Given recent debacles (https://news.ycombinator.com/item?id=36370685), I'm wondering how we can design LLMOps tools for evaluation which both support the use of LLMs as scorers, but also caution users about their results. Are you thinking similarly about this question, or seen usability testing which points to over-trust in 'auto-evaluators' as an emerging problem?
No problem! I guess I will make a plug myself --we've been working on a similar 'prompt engineering' tool, ChainForge (https://github.com/ianarawjo/ChainForge). It's targeted towards slightly different users and use cases than promptfoo --geared more towards early-stage, 'quick-and-dirty' explorations of differences between prompts and models for less experienced programmers, versus the kind of continuous benchmarking and verification testing power that promptfoo offers.
I particularly like promptfoo's support for CI, which I haven't seen anywhere else, and is very important for developers pushing prompts into production (esp since OpenAI keeps updating their models every few months...).
You’re missing the point here. It’s not even getting the LLM’s opinion on evaluating the responses to the prompts (which itself is fraught for some tasks, and benchmarks are known to be limited —even OpenAI admits this, it’s why they made evals). It’s one level abstracted from that. It’s evaluating what the LLM thinks of how well the prompt will do, in purely hypothetical terms. That’s hogwash —different LLMs perform very differently even for the same prompts. Try any tool that lets you compare model responses side-by-side. Unless I see actual use cases, this is yet another iteration of overtrusting AI.
Here is what HN was talking about, nearly three months ago -the exact same type of ‘auto-prompt-gen’ tool: https://news.ycombinator.com/item?id=35660751
This tool doesn’t benchmark based on how a model actually responds to the generated prompts. Instead, it trusts GPT4 to rank prompts simply in terms of how well it imagines they will perform head-to-head. Thus, there’s no way to tell if the chosen ‘best prompt’ actually is the best, because there’s no ground truth against actual responses.
Why is this so popular, then (more popular than promptfoo, which I think is a much better tool in the same vein)? AI devs seem enamored with the idea of LLMs evaluating LLMs —everything is ‘auto-‘ this and that. They’re in for a rude awakening. The truth is, there are no shortcuts to evaluating performance in real world applications.
Nth time this week someone repackages a call to GPT4 as a way to improve/evaluate LLM outputs. Guys, just stop.
Looks a bit like snakeoil to me. A lot of companies now spinning up simple demos with opaque backends, making huge claims they’ve solved X hard problem for/with AI, then saying “trust us” and “join our waitlist” without hard details or facts to show for it. If you could detect hallucinations/biases etc that easily, don’t you think OpenAI would’ve worked on something like this?
Honestly, I wouldn’t consider these domains “foundational” for undergrads —data structures and algorithms and programming and complexity, but database SQL queries? Maybe the issue is what you’re calling foundational here.
We did this with notational programming: https://www.youtube.com/watch?v=wpbOyBqLORU
This paper talks about the history behind how "writing" code came to mean "typing" it. Originally, the very first programming notations were handwritten. Interestingly, they weren't called "languages" either, until FORTRAN introduced the typewriter as semi-direct input. See here: https://dl.acm.org/doi/abs/10.1145/3313831.3376731
The issue is that it's hard to determine who is considered "black" and who is not since race is a made-up fiction constructed by white supremacists that is now so engrained in American society we find it difficult if not impossible to reflect on it. Are we using the Fitzpatrick skin color scale? Are we continuing the hypodescent rule (an algorithm, btw)?
Maybe we just outlaw face recognition in criminal justice entirely.
Implicit in this argument is an egocentric and individualistic viewpoint in universalizing how the author learned as "the future" a.k.a. the way everyone should learn. I don't fault the author for having this, many people do. I wish the author reflected on the role of social factors in education. Learning in this post is something that's thought of as individualized, which is something that educators and cultural psychologists have pushed back on for decades. It totally ignores the role of teachers.
Consider: "AI will make humans vastly more effective by automating tedious tasks." Sure, but which humans? If games are indeed "the future," let's pay teachers even less than we do now, or fire them --that will be the response. That's in fact what Graham says --"I had examples to work from, but no teachers or classes." Let's be real here: Graham's father was a nuclear physicist, he likely grew up rich, and he attended Ivy League schools. Maybe he "didn't need teachers" because, I don't know, his home life was very supportive and he knew how to engage in the dominant, elite American culture? I like Graham here, but c'mon man.