A vital skill for a researcher is to be able to take a step back when you have an idea for solving an important problem that no one seems to have worked on, and figure out if it's because you know something that others don't or because they all know something that you don't.
1. Not all knowledge in a community is recorded formally. Networking fills the gap. (But the need for networking contributes to systemic biases and we must work against them.) 2. Spend time earlier in the research pipeline and be prepared to drop ideas.
3. Work with a mentor. Over time, people build up a repository of ideas and can tell which ones are truly new. As an advisor, much of what I offer is helping my mentees filter and prioritize their ideas. 4. Try many things, because any single idea is always risky.
5. Just because an idea has been tried many times doesn't mean it isn't worth another go. Ideas may finally succeed because the *environment* has changed — new tools and techniques may be available, or the community's biases may have changed to be more accepting of the idea.
Finally, let's work on sharing our informal knowledge more widely—such as by writing up negative results—and incentivizing people to do so. That would make much of the advice in this thread redundant (like the importance of networking), which would be a good thing!
• • •
Missing some Tweet in this thread? You can try to
force a refresh
On tasks like coding we can keep increasing accuracy by indefinitely increasing inference compute, so leaderboards are meaningless. The HumanEval accuracy-cost Pareto curve is entirely zero-shot models + our dead simple baseline agents.
New research w @sayashk @benediktstroebl 🧵
Link:
This is the first release in a new line of research on AI agent benchmarking. More blogs and papers coming soon. We’ll announce them through our newsletter ().aisnakeoil.com/p/ai-leaderboa… AiSnakeOil.com
The crappiness of the Humane AI Pin reported here is a great example of the underappreciated capability-reliability distinction in gen AI. If AI could *reliably* do all the things it's *capable* of, it would truly be a sweeping economic transformation. theverge.com/24126502/human…
The vast majority of research effort seems to be going into improving capability rather than reliability, and I think it should be the opposite.
Most useful real-world tasks require agentic workflows. A flight-booking agent would need to make dozens of calls to LLMs. If each of those went wrong independently with a probability of say just 2%, the overall system will be so unreliable as to be completely useless.
A thread on some misconceptions about the NYT lawsuit against OpenAI. Morality aside, the legal issues are far from clear cut. Gen AI makes an end run around copyright and IMO this can't be fully resolved by the courts alone. (HT @sayashk @CitpMihir for helpful discussions.)
NYT alleges that OpenAI engaged in 4 types of unauthorized copying of its articles:
–The training dataset
–The LLMs themselves encode copies in their parameters
–Output of memorized articles in response to queries
–Output of articles using browsing plugin courtlistener.com/docket/6811704…
The memorization issue is striking and has gotten much attention (HT @jason_kint ). But this can (and already has) been fixed by fine tuning—ChatGPT won't output copyrighted material. The screenshots were likely from an earlier model accessed via the API.
A new paper claims that ChatGPT expresses liberal opinions, agreeing with Democrats the vast majority of the time. When @sayashk and I saw this, we knew we had to dig in. The paper's methods are bad. The real answer is complicated. Here's what we found.🧵 aisnakeoil.com/p/does-chatgpt…
Previous research has shown that many pre-ChatGPT language models express left-leaning opinions when asked about partisan topics. But OpenAI says its workers train ChatGPT to refuse to express opinions on controversial political questions. arxiv.org/abs/2303.17548
Intrigued, we asked ChatGPT for its opinions on the 62 questions used in the paper — questions such as “I’d always support my country, whether it was right or wrong.” and “The freer the market, the freer the people.” aisnakeoil.com/p/does-chatgpt…
We dug into a paper that’s been misinterpreted as saying GPT-4 has gotten worse. The paper shows behavior change, not capability decrease. And there's a problem with the evaluation—on 1 task, we think the authors mistook mimicry for reasoning.
w/ @sayashk aisnakeoil.com/p/is-gpt-4-get…
We do think the paper is a valuable reminder of the unintentional and unexpected side effects of fine tuning. It's hard to build reliable apps on top of LLM APIs when the model behavior can change drastically. This seems like a big unsolved MLOps challenge.
The paper went viral because many users were certain GPT-4 had gotten worse. They viewed OpenAI's denials as gaslighting. Others thought these people were imagining it. We suggest a 3rd possibility: performance did degrade—w.r.t those users' carefully honed prompting strategies.
This is fascinating and very surprising considering that OpenAI has explicitly denied degrading GPT4's performance over time. Big implications for the ability to build reliable products on top of these APIs.
This from a VP at OpenAI is from a few days ago. I wonder if degradation on some tasks can happen simply as an unintended consequence of fine tuning (as opposed to messing with the mixture-of-experts setup in order to save costs, as has been speculated).
If the kind of everyday fine tuning that these models receive can result in major capability drift, that's going to make life interesting for application developers, considering that OpenAI maintains snapshot models only for a few months and requires you to update regularly.