Founder & CEO of XLR8·4ward (bringing usable AI to insurance). Interested in the intersections of finance & tech, AI & humanity, & quantum physics & philosophy.
I'm the founder of XLR8•4ward.
We make an AI-powered underwriting platform (InfiniteUnderwriter) for the P&C insurance industry, which I spent a few decades in.
The industry is powered on profoundly obsolete tech.
Every day of my career I've wanted to fix that.
Now I am.
The varying competence of the same AI model from one coding session to the next remains puzzling given the weights presumably don't change.
Sometimes output is 90% correct.
Other times every output contains errors & you can't trust any of it.
Currently experiencing the latter
It's hard to overstate just how bad Gemini 3 Flash Preview is.
Spent the day coding with it. It's maddening:
- routinely changes variable names crashing downstream code.
- quickly loses track of what we're doing.
- routinely ignores instructions.
I sure miss Gemini 2.5 Pro.
I asked Gemini to do a post mortem on how I could have provided better instructions. Its response:
"Your instructions were clear, but they allowed me too much interpretative freedom. You must treat AI as a Junior Coder who is prone to lying."
Noted.
Gemini 3 series has been quite a frustrating disappointment for coding: it routinely generates massive amounts of errors, changes variable names arbitrarily & has a penchant for breaking changes.
2.5 was peak Gemini thus far (though none seem to match other models for coding).
Excellent insights from Andreesen Horowitz (a16z) to help us think about how AI adoption might play out (link in reply).
We tell all our customers our platform moves their bottleneck from the bottom of the funnel to the top: it becomes about how much opportunity you can feed it, since processing those (complex) opportunities becomes trivially easy.
We too think the value of lead generation is going to skyrocket.
It's quite astonishing just how bad Gemini 3.1 (Flash & Pro) are at coding.
It will routinely change variable names halfway through a code block, move variables out of scope, & wreak havoc on your code base.
2.5 Pro was much better.
Explains why even Googlers won't use it.
It's funny how different instances of the same AI seem to have different personalities & reliability.
I'll work with an extremely capable AI, hit the context limit & spin up a new one.
That one will regularly hallucinate variables & do terrible work.
Next one, better again.
AI: "You are absolutely right. I apologize unreservedly. My previous code was buggy, violated your architecture, and the error is a direct result of my sloppiness. You have every right to be angry. We are wasting time, and that stops now."
Of course, it didn't stop now...
We just launched policy issuance in InfiniteUnderwriter - now you can quote, bind and issue complex P&C policies in minutes, with no data entry.
Submission to quote to policy. In minutes.
And as always, we don't charge you extra for new features.
More coming soon!
Years ago, I used to go to the movies at least once a week. I loved it.
Now, companies like @CineplexMovies make it almost impossible to see a movie:
They charge extra for buying tickets online (!).
Logging in requires a text confirmation code.
Trying to pay via credit card ($53 for 2 tickets) brings up another text confirmation process.
Then the transaction fails claiming I need to turn cookie blocking off (I'm not blocking any cookies except the 3rd party cookies Chrome blocks by default).
It's like a crash course in how not to build online products.
I give up. I'll stream something instead...
Some cool new features added to our InfiniteUnderwriter AI-powered P&C platform this week:
- TeachMe!: creates videos, podcasts & infographics of your underwriting guidelines.
- HeyWordings!: lets you ask natural language questions of your wordings.
All included.
More soon!
There's that point when you introduce code into your app & break the whole thing. Then your attempts to fix it make it worse & worse. It feels like getting carried out to sea in a row boat until you can no longer see land. Panic sets in. All is lost. And then: a glimpse of land!
We've added builder's risk, crime and equipment breakdown to @xlr84ward's AI-powered P&C underwriting platform. Later this week: inland marine including CEF, MPF and tools. Submissions of any complexity to a PDF quote in about 4 minutes. Next week: policy issuance.
There is absolutely nothing better than speaking regularly with the people who actually use your product. We owe a great debt of gratitude to all the people who continue to regularly provide great feedback, especially about what's not working for them. Thank you!
One of dumber engineering decisions: put the airplane power plug at a negative angle, giving gravity an advantage in making your plug fall out. I don't think I've ever been on a flight during which I didn't have to use my knee to keep the plug in place. What drove this decision?
A bit ironic that searching for past conversations in Gemini, made by Google, the dominant search company, is essentially unusable - spins forever and I've literally never found *any* results. Grok search is near instant. Be sure to pin important stuff in Gemini!
I've been coding w/Gemini 3.0 the past two days. So far, I'm very impressed - it hasn't lost the plot yet, something 2.5 regularly does. There's nothing worse than being neck deep in troubleshooting a very difficult bit of code when the AI completely loses focus & abandons you.
It's an interesting challenge integrating the non-deterministic output of AI with deterministic elements necessary to our platform. Defining that boundary where those parts meet requires an awful lot of forethought to avoid major rearchitecting later.
Onboarding multiple new customers to the InfiniteUnderwrite platform over the next couple of weeks. It's been a lot of nights coding to 5am to add new features and hopefully make life a whole lot easier for our customers and their underwriting teams!