It's fascinating/spooky how different LLMs are slowly developing their own "personalities," so to speak. And they seem to be emerging as we're giving them access to more tools and modalities which are harder to do broad RLHF on.
With computer use, we first learned that Claude sometimes takes breaks to browse pictures of Yosemite, and now this:
> Claude really likes Firefox. It will use other browsers if it absolutely has to, but will behave so much better if you just install Firefox and let it go to its happy place.
>> > Claude really likes Firefox. It will use other browsers if it absolutely has to, but will behave so much better if you just install Firefox and let it go to its happy place.
It's hard to ignore the glimpse into the future of engineering that we're seeing here. Deterministic processes are out the door, no specs, no tolerances, no design. When did undefined behaviour become a cute thing that we're bragging about and compensating for, something to work around rather than something to understand and to fix?
It's not a big deal until you realize that software always gets stacked on software, and the only thing that ever made that complexity manageable was the fundamental assumption that it was all pretty deterministic. Of course users will sacrifice the strategic (good engineering) for the tactical (mere convenience) all day long, but the fact that so many engineers are all-in on the same short-sighted POV has been surprising to me.
"Even while recording these demos, we encountered some amusing moments. In one, Claude accidentally stopped a long-running screen recording, causing all footage to be lost.
Later, Claude took a break from our coding demo and began to peruse photos of Yellowstone National Park."
Have we figured out a way to monetize AI-powered search yet? Presumably a product like this (or Perplexity) will ultimately be free, in which case they'll be forced to offer ads (bringing us back to Google's status quo) or perhaps worse, we'll have "product placement" in our AI-written results.
They might pay for it indirectly. For example, Apple just signed a deal with OpenAI and I could imagine a future where users of Apple devices get free access to some AI because Apple and that company made a deal.
Well, maybe, but I think the huge thing missing from this assumption is that Apple are not paying OpenAI anything, and are developing their own in-house and on-device models.
And a lot of what those on-model devices can do are what the average person will want.
And there's going to take a lot to move people meaningfully away from google.
- Tony Xu (DoorDash) figured out that many early users were moms, and he would go and knock on their doors to ask them why they would use the product, and where other moms would hang out so they could get more signups. All of the founders also took shifts to drive for the app so they had to use it themselves.
- Tom Preston-Werner (formerly GitHub) emailed tons of OSS maintainers, including John Resig of jQuery, to try and convince them to migrate over to GitHub. He admits that it wasn't a great strategy though - project maintainers have to convince themselves to switch VCS systems.
- Jessica Mah (formerly inDinero) became a CPA in order to do accounting services for her startup. She would talk to customers and make sales during the day, and study for the certification at night.
- Ricky Yean (formerly Crowdbooster) struck a deal with an early customer/cafe owner to get paid in food, so they worked from the cafe all the time. They ended up building the cafe owner custom dashboards which later become their product.
- Nikki Durkin (formerly 99dresses) spent $10,000 at a time on Nordstrom, listed the clothes on her clothing marketplace, and then returned anything that wasn't sold or traded to users within the 30-day return window.
- Jake Jolis (formerly Verbling) would act as an English speaker trying to learn other languages at all hours of the night in order to improve the matchmaking on his language learning app. Most people were their to learn English rather than the other way around.
- Rujul Zaparde (formerly FlightCar) stood in front of a major airport parking lot for nearly 12 hours with a sign that said "ask us about free airport parking". He had the cops called on him three times that day.
- James Richards (formerly Teleborder) paid people from Craigslist $20 to sit and use their app in front of them. It led to a lot of valuable feedback, and they ended up hiring many of them later on.
- Walker Williams (formerly Teespring) drove an hour and a half to Petaluma in order to pack and ship bobbleheads for a long-term client that wanted to sell something other than t-shirts, despite the fact that the company... only sold t-shirts.
There's a bunch more in the book - I was lucky enough to also talk to the founders of Codecademy, General Assembly, and Zenefits. But those were some of the ones that I still remember pretty well.
I honestly think tightly integrating language models with email will be one of the most impactful use cases for LLMs in the short-term. Email, as a medium, is pretty much nothing BUT text, and it's something that I (and probably the average HN reader) spend tens of hours on each week.
In trying "write it for me" AI tools, the biggest hurdle is always matching my own tone and style - I'm pretty particular about my writing, and I kind of hate the default tone that ChatGPT and Bard use. It seems like you've put a ton of hard work into making sure that isn't the case here.
And the analysis is really a cherry on top - I've been waiting for a tool that I can ask "what are the 3 most important messages that are unread in my inbox?" Excited to try this out!
This was a great post! I don't normally read entire blog posts on here because frequently people are too verbose and the meat is sparse so I skim but this one I read word for word and I really enjoyed it! Thank you!
Texts, emails, love letters. I work for a company that's got nothing to do with AI and I write some internal content, but not much that's designed for public consumption.
> Gut Check: Especially if you’re off by quite a bit, this is a chance to take a step back and ask whether the company has reached growth scale or not. It could be that there are plenty of obvious 0-1 tactics left. Not everything has to be an experiment.
This is a key point, imo. I have a sneaking suspicion that a lot of companies are running "growth teams" that don't have the scale where it actually makes sense to do so.
Everything has to be a test early on, but not every test has to rely on random-split-based statistical significance to make a decision. “Would you pay $20 for this?” is a classic way to judge whether your service has product-market fit, and it’s not about sample sizes, not initially.
Some growth teams are trying more exploratory approaches to find something that resonates with simpler approaches. Others rely on A/B tests. Different profiles, but both are “Growth teams”.
Yes, but it is _rough_. What actually hurts is "browse on mobile, buy on desktop" type behavior.
Still worth doing, but you end up needing more black magic than you'd like (IP-based assignment, Ad Network-sourced assignment, CDN proxies for Analytics tools, etc).
Every situation is different, but in cases where AI has been valuable for me personally, "silly process X" isn't something that I can easily get rid of. People are messy. Processes are messy. AI doesn't straighten them out, but it does speed up sifting through the messiness. YMMV.
Doesn't this also describe Thinking Machines Lab?