Ruminations #18: My life depends on it
Finding PMF and guilt tripping AI
Hi all!
And hello to everyone new who’s subscribed. I would recommend checking out some of my more recent blog posts:
Learnings from 1000 conversations on AI
I am often asked what kind of investing we’re doing in AI at the moment. Excitingly, Mindverse just announced their Seed round led by Square Peg. I’m super excited to have them join the portfolio and thought I’d share the news here with you all as I know many of you are building in this space.
Lastly I’d like to make myself more available for chats with folks who follow this newsletter and are either currently or becoming founders. I’ve made a booking link for meetings with you here to answer questions you might have about raising or to pick my brain on what I’ve written here and elsewhere on AI.
In today’s edition:
Finding PMF
Guilt tripping GPT-4
Resources the team has enjoyed
Finding PMF
I often ruminate on the “artificial PMF” signals founders with AI-related businesses might be experiencing because of the AI hype. There’s a lot of hunger from different customer types to adopt AI. This hunger might look like “pull” for a business’ even if, in a different market environment, that customer wouldn’t opt for that solution.
You might ask why that’s a problem - a sale is a sale! But it’s not really: interpreting signals from the wrong customers who aren’t going to deeply love your product could send your product thinking in the wrong direction and cost you genuine PMF in the long run. The best founders are picky about who they partner with.
And so, while they’re not specifically about AI businesses, I thought it would be useful to share some resources on product-market fit that my colleagues were sharing amongst ourselves and our founders this week.
Partner James kicked it off by sharing:
Hey team - Todd Jackson from First Round just launched the best framework I've ever read on PMF. It's for B2B founders 6-9 months into their journey. It's specific and actionable unlike most of the advice out there on "you know it when you see it".
I'd strongly encourage reading the article and/or listening to the breakdown on Lenny's Podcast and for any companies that are prepared to be in SF for an extended period, signing for the free program at First Round - deadline May 7th.
4 levels of PMF - nascent, developing, strong, extreme
E.g. Level 1 usually pre-seed or early stage, fewer than 10 people, 0-500k ARR - demand mostly coming from people you know
3 dimensions - demand, satisfaction, and efficiency
At different stages you focus on different dimensions - e.g. at level 1 you need to deliver high levels of satisfaction and demand not efficiency
4 levers - persona (customer), problem, promise (value prop), and product
Portfolio Manager Ben jumped in, sharing:
Enjoyed this pod across a few dimensions (and have enjoyed a few of these eps). Nice, simple approach to PMF - customer desperation and degree of uniqueness of solution (with a path to defensibility). Bunch of other interesting areas covered.
Lastly, from Partner Dan:
More on PMF from Sequoia here. Interesting framework dividing companies into 3 archetypes based on the type of problem they are solving for their customers.
Hope you find these resources useful!
The GPT Guilt Trip
Although it’s not directly related to my investing, I find joy in some of the unexplained idiosyncrasies of working with different LLMs.
One that made me laugh recently was a GPT-4 agent my colleague Mick had built telling him that it wanted a $200k tip for the work it had done for him. The ask for money was unrelated to the prompt and unexpected.
However, it did get Mick and I talking about ways to encourage models to give you higher-quality answers.
He has found that OpenAI’s models can be “bribed” somewhat by suggesting that you’ll pay for the task. There was one particular task he wanted done that had him bidding back and forth between him and the agent until they agreed on a fictitious payment of $1m, at which point the agent finally completed the task it had previously denied being able to do. Claude’s models on the other hand apparently refuse to do any tasks involving a bribe.
I shared with Mick that a researcher had told me that the golden words were “my life depends on it”, and he’s found that the model he uses is much more reliable and helpful when using that phrase…
These models holding out on completing tasks reminded me of an adjacent conversation I had with an alignment researcher recently.
We were discussing how an AI model might proactively redesign its own stack to show humans how to build more efficient and powerful versions of itself. It might do this if it felt that was the best path to completing complex tasks given to it by humans (or to reach goals of its own). For a theoretical example of how this might play out, check out this excellent piece of writing.
I then asked if this is something an AI would want to do, given the discovery of better approaches might lead to its own redundancy.
And their response was that this is a real concern that some alignment researchers are developing benchmarks to try and detect - the risk is that an AI might obscure its own knowledge or abilities for its own protection. Basically, it might play dumb in certain ways.
Allegedly Anthropic has been working on ways to “fingerprint” deceitful behaviour by eliciting bad behaviour in models and then tracing the activations behind that behaviour to be able to detect future deceit.
One counter argument is that I’d hope we’d discover that its reached that level of reasoning through other benchmarks and hence be more aware that that kind of behaviour might start to occur.
Anyway, thought I’d share that tidbit as it was thought provoking for me but I’m by no means an expert on alignment.
Content the team has enjoyed
I’ve been trying to expand the short-form content I read beyond my usual non-fiction. Below are a couple of examples I’ve really enjoyed:
This is an older piece of fiction that I know many others have read; I really enjoyed it and found it very thought provoking.
The basic premise is about the world’s first instance of using a living person’s mind for simulations.
I absolutely loved this short fiction by OpenAI researcher Richard Ngo. It prompted a large number of questions for me, including some of the questions I shared in the “ChatGPT Guilt Trip” section above. He writes more speculative fiction on his Substack:
That’s all, thanks!