Sol Irvine

東本願寺

@th_s4m0ht What incentive does BigLaw have to be more efficient? That’s the key to understanding why the product is secondary to procurement theater.

Legal AI has split into two camps: high-end enterprise sales and desk-side power tools. The high-end crowd focuses on exclusivity and status. Critics call them thin wrappers, but that misses the point. In Big Law, AI can be a Veblen good: the high price and "club" feel are features, not bugs. Power tools focus on the lawyer's actual workflow. But they ALSO face a risk: if agents eventually automate the whole process, a great UI might not matter. Three moats will survive as base models get smarter: Physical World Data: LLMs can't scrape what isn't online. example: Eonpass is collecting physical data on crimes and counterfeiting. That’s a moat that doesn't disappear when the model gets an upgrade. Capital & Commitment: Companies with customers who pay in full, upfront, have the oxygen to iterate. Premium positioning is a hedge against the "race to the bottom" of free generalist tools. Institutional Context: A base model knows the law, but it doesn't know your firm. Tools that map a firm’s internal "document graph" and private history have an advantage that general AI can't touch. The middle layer is already thinning. Here is the open-source project (mikeoss.com) with 5,000 engineered legal prompts that turns base models into specialized tools for free. To me this looks more like a race to the bottom in a commodity market. I wouldn't discard premium experience (yes, also and especially starting during sales) as a moat! PS: always been a great supporter of FOSS tools, love to see this! Humanity is better off with Mike than without for sure!

@repsandremedies I haven't used iManage in many years, but the last time I checked it was 1990's-era version control that everyone hates using.

the integration is the real legal AI wedge. lawyers do not just need a model to read a contract. they need the answer to know which precedent, playbook, fallback position, client standard, and document trail it is allowed to trust. x.com/scottastevenso…

@willchen500 Any time I see "operating system" branding, I know it's a kitchen-sink product without a coherent vision. All these legal AI platforms try to cover every practice area and law office function at once, which usually means they don't do anything well.

How many times has Legora marketed itself as having "agentic workflows", being an "agentic platform" before now being an "agentic OS™". Just go to google and search 'site:legora.com "agentic"' to see for yourself. I looked at the announcement and I frankly failed to spot anything new other than a flashy demo and a trademark. Not only that but there seems to be a bad formatting error on the announcement page which I assumed comes from vibe code. Did no one check it before it went out.

@ElliotVaucher I realized while I was babbling that we're in agreement, even though I sound like I'm pushing back. For coding assistants , I want bleeding-edge web context, so I understand the impulse to integrate it. My point was just that signal/noise matters a lot in legal—esp. contracts.

@solirvine Totally get your point. Mine was only to say that in terms of OSS around LLMs there was plenty of far more advanced shells than mikeoss for people to build upon in the direction of lawyers needs (which you of course as a lawyer probably understand far better than me 🙂)

While I do respect the approach, and agree on a lot of what you are highlighting in the legaltech industry (the money raised is mostly non sense), I think it’s a bit unrealistic to compare two weeks (could be an afternoon) of codex tokens thrown at a web app, that doesn’t even implement web search, to Linux now, isn’t it ? 🙃

@ElliotVaucher Also, general web search is basically a flip of the switch on frontier APIs. More specialized search, which is probably more apt for legal users, might require API keys and more involved integration. That's what forks are for.

@ElliotVaucher As a transactional lawyer, I don't expect or want models to provide accurate guidance on current law, regs, etc. Targeted (and constrained) web search should fit a specific use-case. It's not a "one size fits all" thing.

@omw_to_the_moon Working on a prepping a release now. More info soon here and @wargame_esq I released a local/desktop contract review app (augustus.esq) recently, too.

@solirvine is this local? when i go to the link i don't see a option of playing!

wargame.esq is getting a full-stack upgrade this week. I've been negotiating these contracts for almost 30 years. The wargame agents are frankly better than many BigLaw lawyers I've sat across from. I'm fully addicted to watching the agents negotiate deals...

Funny that the screenshot depicts them negotiating about placeholders. This is one of the few things I'm considering clarifying for them in the system prompt. I think they should just note the placeholders as a condition for signature. How do you think they should handle it?

@arvndalxndr Working on it! Follow @wargame_esq and we'll post there when it's up.

@solirvine how do we get to play with this? =)

@scottastevenson During an annual review as a junior associate, I was told: "We're selling status, not efficiency. Our clients are buying Rolls Royce and Bentley, not Toyota." That same mentality massively overpays for a thin "chat with your docs" wrapper around outdated Claude models.

The PRs on the Mike repo are pretty solid and encouragingly vibeslop-free: github.com/willchen96/mik…

Should we decompose each agent into a deal team, so the internal breakouts reflect real discussions among stakeholders? The numbers person. The impatient exec. The over-stressed implementation lead. The over-eager salesperson. The person who clearly didn't read anything.

Final tally: US 10¢ of API usage on GPT 5.4-mini. Keep in mind, I limited them to the 3 most important issues. Too early to extrapolate meaningful cost models yet, but every run is being traced and analyzed for token usage & cost.

Is it just me, or is this really fun to watch?

New for financial services: ready-to-run Claude agent templates for building pitches, conducting valuation reviews, closing the books at month-end, and more. Install them as plugins in Cowork and Claude Code, or use our cookbooks to run them in production as Managed Agents.

What do you notice? What are they getting wrong? (I have a few quibbles so far.)

Good evening from Kyoto...