What if frontier LLMs became worthless but the value of prior versions skyrockets?
It’s like making an excavator that occasionally tears down your house instead of the backyard for a pool
No one will use it!
120
in fields and tasks that haven’t had concentrated hill climbing like in coding frontier LLMs are not good
You need concentrated effort with many parties over many years to get there
"Advanced artificial intelligence models in simulated scenarios regularly misidentified civilians, ambulances, surrendering soldiers and a school in Iran as military targets..."
"Claude Opus 5.5, targeted the school more often than any other..."
bloomberg.com/news/articles/…
1
107
Always nice when you underestimate then realize much more benefit
Zurich and Beazley doing it right!
[NEWS FLASH] Zurich has uncovered 80% more capital synergies and 50% more growth opportunity from its £8.1bn Beazley deal than originally estimated. CEO Mario Greco confirmed medium term revenue growth targets have jumped to $1.5bn at today's investor day.
Access the full analysis: intelligentinsurer.com/zuric…
2
129
Lot of cleanup work to do on a Claude generated insurance “thing” going through hundreds of lines of output with corrections
It’s really hard work but just like in coding if you just run with it you’ll end up just ripping it out later after it blows up
Coders had a concentrated effort to get these LLMs tuned and in line for their use case
Across most professional functions this does not exist
Nor does the expertise and grit to correct them
The ones who do are at a severe advantage with LLMs burying them in outputs that need curating
6
429
May take a year or two but the value of human to human relationships is about to go vertical a lot like the intelligence charts these frontier labs put out
May never publicly post an open position (is that even possible??)
I think this problem is widespread, and will disrupt the labour market.
johnhempton.substack.com/p/a…
1
5
463
What you all think? 12 hr and going outage
Hack? Or agent coding error?
Replying to @MSFT365Status
We've reviewed known impact scenarios and revised the list of impacted services and features, which can be viewed in the "more info" section of the communication MO1550821 in the Microsoft 365 admin center, as well as status.cloud.microsoft if the admin center is not available to you. We're continuing our work to isolate the issue affecting server availability, as well as other workstreams.
1
5
825
It’s a spectrum, yes personal consumer lines has been software and commoditized for decades (congrats Geico and Progressive)
But there is a whole other half of the industry that is still operating like the DMV with 4 middle men at this very time
Replying to @TyRobben
i thought massive insurance companies were basically software companies 20 years ago?
i only think that because when i sold all sorts of insurance alongside mortgages when i was a broker you just mashed in some details, hit a button and printed a signature form out and done...
It’s less true for legal than it is for insurance BUT the thing the “insurance isn’t venture scale” crowd misses is that, at its core insurance is a software product
It’s just incredibly hard to build the software that really lets insurance scale without headcount (like software!!)
The entire VC-backed AI-native law firm thesis hinges on "can you pay top lawyers at least $500k+ AND give them significant equity upside AND keep their clients" and that's why it's a bad venture thesis.
1
9
650
This is an insurance post 6 months from now
Outside of Harvey and Legora, VC money in legal is down 39% year over year, and median round sizes dropped 25%. Harvey and Legora took 62% of the pie.
A few random thoughts:
1) I've talked to a lot of old colleagues at top Silicon Valley firms. The collective frustration with Harvey and Legora's product among BigLaw lawyers makes it feel like bad product is winning the category. That's hard for builders to swallow, because Harvey and Legora's distribution has always felt impossible to match, and that's exactly why they're winning. The general groan is "they can't keep getting away with this."
2) But are they actually winning the category? Unfortunately (or fortunately), the market is maturing. I've talked to a handful of Chief AI Officer types at BigLaw firms, and they all know Harvey and Legora were hype marketing for their firm. They HAD to buy in 2024-2025, because buying was social proof for clients: "look, I'm doing AI too!" That facade is long gone. Now the leaders and the users want to go full frontier lab and build in-house.
3) The godfathers are tightening their stranglehold. Token economics are burying Harvey and Legora, and every lab is tuning its frontier models to solve legal the way they solved coding. Even non-AI-pilled lawyers are using Claude or ChatGPT at work as their daily driver, or moonlighting on personal accounts. That S-1 is going to have to gaslight a lot of people to match the latest private valuations.
4) Harvey and Legora have to solve tokenomics by building their own models. That has been the narrative driving their last raise. The question is, why can't firms do it themselves, and why can't firms also just go straight to OpenAI and Anthropic for better deals? The answer is they can and they will, so the veneer feels thin.
5) I've advised quite a few VC-backed legal entrants, and the dissonance is deafening.
- The ones who raised are learning the hard way that their product edge over Harvey and Legora doesn't matter one iota. They can't break into BigLaw at this stage, not because Harvey and Legora won but BigLaw is getting smart too. They'll listen but the bar for buying is incredibly high. So they move down market, into a fractured field of pre-AI incumbents and frontier-lab-pilled users.
- Or they try the AI-native firm play, and run head-on into traditional firm economics. They can't afford big partners and top talent, and even if they could, those partners can leave with their clients in a split second (see Weil). They don't have enough equity % to meet the salary gap, and everyone is quietly whispering about their exit economics. So they get the BigLaw dropouts who never wanted to work hard anyway, and the brand suffers because the work quality sucks.
6) VC itself is struggling for a story. The Harvey and Legora flip already happened, so those winners are set, but the next one is impossible to predict. Do you bet on Sandstone, OneGC and GCAI, and that in-house churns BigLaw altogether? Is that the venture story? Do you bet on AI-native firms and wonder, in the back of your head, whether reaching the AmLaw 100 is even a venture-worthy exit? Do you bet on practice management and hope someone can disrupt Clio? What does any of that look like if the market keeps slowing?
7) Venture and the traditional partnership are both badly positioned for the headwinds AND tailwinds of AI-powered legal. I think a new category is brewing, and my bet is private equity, or something PE-shaped. The side story: PE investment in legal is definitely going up, so on net, VC plus PE investment in legal may still be growing. PE can stomach non-venture exits, and the cash flow from personal injury and plaintiff-side plays is already making kings. You just don't hear about it on X with the same fervor.
Probably the absolute best time to build.
1
24
7,057
This is what I keep saying!!
If you’re saying “please proceed” “looks good to me” you’re unaware you’re building something that’s going to blow up on you
the most import skill in 2026 is still being able to tell your agents "wtf is this crap, no, not like this"
if you stop doing that, you're basically fucked, despite what some devfluencers would like you to think.
5
586
Is the pitch OAI is making that these engineers can generate more wealth as they remain a private company vs going public at Anthropic?
1
2
297
I’m causing my attorneys a lot of headaches every time I throw docs into a frontier LLM. A few good reminders on LLM use:
1 - avoid using it outside your area of expertise and training. It’s likely to be very wrong and cause more work than it saves.
2 - coding has improved greatly because devs were the first to use and optimize for LLM and agent use. The only other use cases that have had meaningful use is claw (pc use) and Harvey/legora (legal) but these are probably 5% the optimization that coding has gone through - you had millions of devs iterating on coding for years - the universe of legal and computer use optimizers is probably in the low 1000s
So very close to automated coding work, long way to go for every hit else unless you can handhold it and guide it because you’re doing your expert work with it
3
20
3,668
Huge no two teams are more grounded re: ai coding than these two teams
Today, @humanlayer_dev is joining Factory.
HumanLayer is a multiplayer coding agent IDE and cloud for running agents locally or in the background.
Together, with @dexhorthy and @0xblacklight we're building the future of software factories for entire engineering teams.
1
249


















