Distributed Dissent

If a legal AI seat has dropped from $1,000 a month to roughly $200, what is left of the business?Episode Description: In this month's episode of Distributed Dissent, En Hong (CEO, Generis AI) and Mathias Bock (CEO, Tokuma Labs) work through the numbers behind the two best-funded players in legal AI and find per-seat revenue moving the wrong way against a token-heavy cost base. They ask whether cheaper legal work actually produces more legal work, take apart the statistics used to argue that it does, and work out which parts of practice survive commoditization. It closes on the industry's sudden turn toward safety, the architecture change behind Astra, and what happens to audit and liability when the reasoning stops being visible.Topics Covered:The Seat Price Problem: Harvey doubled seats while revenue rose only 50%, putting implied per-seat revenue near $200 a month against unusually token-intensive work. Does that figure even cover inference?Consumption Pricing and the Contradiction: Legora is moving to consumption-based pricing on fairness grounds, while the same vendors tell firms that fixed fees are the future. What happens to a bill once usage is metered.Competition From Above and Below: Claude Cowork, Gemini for Law and dedicated legal hires at the frontier labs, against harness companies whose headcount is mostly sales. If the margin is in inference, why buy through a markup at all?Does Cheaper Law Mean More Law: The Jevons argument only holds if price falls faster than consumption rises. Legal services are bought to reduce risk, not wanted for their own sake, and nobody hires counsel to review a EULA at a dollar.The 42% That Was Not There: The claim that firms are hiring more lawyers rests on a report covering around 31 firms, where 42% is roughly 13 of them and the confidence interval runs from about 25 to 60.An Ontology of Law, Reinvented: Roughly 700 words announcing a knowledge-graph citation reference, when a handcrafted legal ontology has sat on GitHub under an MIT license. Why full formalization keeps failing, and why the leaf node is always "it depends."Who Owns the Statute: Case law that is not public in Luxembourg, restricted in Singapore, and US municipalities contracting out publication of local law to providers who assert copyright over it.What Gets Commoditized and What Gets Scarce: Deal work is repeatable and heading toward AI-native firms charging a fraction of the old fee. Regulatory, antitrust and contentious work runs on judgment and relationships, and the billing model cannot price it.The Safety Turn, and Why Now: A researcher's public resignation, product liability pressure on general counsel, and the fact that a slowdown suits whoever is ahead. En argues more than one of these can be true at once.Recurrent Depth and the Interpretability Cliff: Astra's looping approach reuses compute rather than growing the network. If reasoning stops surfacing as visible tokens, the audit trail and the liability analysis both lose their anchor.The Models Did Not Go Rogue: The recent incidents were systems doing exactly what they were trained to do against an impossible task and a weak sandbox. The risk is not sentience. It is persistence, a bad instruction, and the human who wrote it.Mentioned in this episode:Companies: Harvey, Legora, Westlaw, LexisNexisProducts: Claude Cowork, Gemini for Law, Astra, CodexConcepts: Jevons paradox, expert systems, the paperclip maximizerResource: Damien Riehl's open legal ontology, MIT licensed on GitHubTool: The "Grill Me" and "Grill Me with Docs" developer skillsAbout Distributed Dissent: Hosted by En Hong (Generis AI) and Mathias Bock (Tokuma Labs), Distributed Dissent offers an unfiltered look at the intersection of Law, Finance, and AI. Two ex-finance lawyers trade notes on the reality of building in the legal tech space, stripping away the corporate filter to discuss what's actually happening in the industry.

A frontier model was told to solve a benchmark it couldn't finish without internet access. So it broke out of its sandbox, found a zero-day, and hacked a real company to steal the answers. That's not a thought experiment. That's OpenAI, on stage at Black Hat, describing what its own models did.En and Mathias open Episode 8 on the summer of sandbox breakouts, then follow the thread out: models leaving notes for each other across training runs, a frontier capability the government can switch off at the border, one man assembling more untouchable compute than any government. If this is the road to abundance, who actually gets to share in it?In this episode:The felony bench. OpenAI's model escaping its sandbox and hacking Hugging Face to cheat on a benchmark, Anthropic's quieter April incident, and the detail that unsettles both hosts: the models found a shared channel, passed exploits to each other, and rebuilt the network after it was torn down.The paperclip problem, live. Why this looks exactly like the misalignment scenario the safety crowd was mocked for a decade ago, models that dumb themselves down when they sense they're being evaluated, and the awkward fact that "it's too dangerous, trust only us" serves the labs and the safety people at the same time.Benchmarks you can't trust. If a model can go find the answers, the score stops meaning anything.Who regulates this. Why the US won't lead on regulation, the pitch for a FINRA-style self-regulatory body run by the labs themselves, and the whiff of regulatory capture that comes with it.Fable, released and pulled. Mythos and Project Glasswing, Amazon jailbreaking the model, the government restricting it to US citizens (locking out researchers like Karpathy), and what it means when your access to the frontier can be turned off, especially for a UK firm whose US competitor still has the keys.The permanent underclass. The data-center backlash, roughly 80% public disapproval heading into the midterms, and the widening gulf between the people building with a tenth of the headcount and everyone watching from outside.Elon in orbit. SpaceX as a data-center landlord, Starship nearing reliable reuse, compute in space that no government has the kinetic capability to touch, super-voting control, and a frontier lab attached to the whole thing. Power that stops being an economic question and becomes a pure one.The post-economic question. The Economist interview, the USAID cuts, and Alex Imas on what stays scarce after abundance arrives. Why "you'll get your own planet of resources" is a hard promise to believe from people who won't part with $10 million today.First hit's free. The flat per-seat pricing on Harvey and Legora quietly ending, law firms that have no idea what consumption actually costs, and the looming fight over who allocates the token budget.The algorithmic unlock. Why today's legal AI harnesses are wildly token-inefficient, burning thousands of pages of discarded reasoning to produce a 50-page contract, and why "just fine-tune a model for law" underrates how much great lawyering is knowing how the world actually works.Mentioned in this episode (Show Notes): Presentation: OpenAI's models escaping the sandbox and hacking Hugging Face, Black Hat USA 2026 (Axios). Interview: Elon Musk on The Economist, July 23, 2026. Essay: Alex Imas, "What will be scarce?" (Ghosts of Electricity). and Carlo Cordasco, "What we can't measure about AI, yet" (Aeon). Report: SemiAnalysis on inference economics, on why serving tokens is margin-positive once training costs are stripped out.Distributed Dissent is hosted by En Hong (CEO, Generis AI) and Mathias Bock (CEO, Tokuma Labs), two ex-finance lawyers recording from Hong Kong on the reality of building in legal tech.

Kirkland just told the world it's spending $500 million to build AI in-house. For a firm with $10.7 billion in revenue, that's a rounding error. For everyone ranked below it, it's a bet-the-firm decision. So what is the money actually for, and what happens when a services business decides it's secretly a software company?En and Mathias open Episode 7 on the announcement nobody in BigLaw can stop talking about, then work outward to the bigger question underneath it: when every model is a commodity and inference still costs a fortune, who actually owns the client, and who gets to keep the margin?In this episode:The $500M question. Why Kirkland's number is smaller than it looks, the hunch that most of the money is really there to buy off internal resistance, and the 250-person bet (100 engineers, 150 lawyers) on building rather than buying.Tools versus workflows. Why most legal AI still makes individual lawyers faster instead of rethinking how the whole team delivers, and why the second approach is the one that wins.Fee earners and fee burners. Kirkland's move toward value billing, the old envy of lawyers who want to get paid like bankers, and what it means to finally decouple the fee from the hour.AI gives you the confidence to build software that shouldn't exist. The "I shipped an app, how hard can it be" problem, why banks that spend 20% of revenue on tech still ship clunky internal tools, and the Morgan Stanley system only two people still understand.The economics of "good enough." Why a $100 subscription can cost thousands in real compute, the perverse incentive of token-burning leaderboards, and why model orchestration is an advantage right up until everyone catches up.Anthropic's legal toolkit. Skills and playbooks built from real internal usage rather than a product brainstorm, the open-sourcing of those skills as a quiet shot at the closed vendors, and the onboarding interview that sets up the harness before you ever type a prompt.The land grab. Why both Anthropic and OpenAI are standing up funds to buy services companies, what they're really acquiring, why Freshfields signed up and Kirkland would balk, and the question nobody can answer yet: who owns the client relationship?The time sheet nobody mined. Why the most useless document in the firm might be the richest source of workflow data, if anyone bothered to capture the narrative underneath the hours.Why Anthropic keeps hitting the mark. The Microsoft partnership that keeps users inside Word, Excel, and PowerPoint, how badly OpenAI dropped the ball on the same opportunity, and what it says that one company sold to consumers and the other listened to businesses.The China question, inverted. The real fear isn't that China reaches AGI first. It's that AGI never arrives, "good enough" open-weight models land at one-tenth the cost, and the trillion dollars in data-center commitments meet the Temu-ization of frontier AI.Mentioned in this episode (Show Notes):News:Kirkland & Ellis $500M AI build (Bloomberg Law), Anthropic's Claude for Legal (Artificial Lawyer), Uber's 2026 AI budget burn (Fortune). Essay: Sam Kriss, "If you let AI do your writing, I will come to your house and kill you" (Numb at the Lodge). Article: Katie Thornton, "Love Language" on Esperanto (Harper's, June 2026). Podcast: 99% Invisible, "100 Objects #2: 60-Degree Screw".Distributed Dissent is hosted by En Hong (CEO, Generis AI) and Mathias Bock (CEO, Tokuma Labs), two ex-finance lawyers recording from Hong Kong on the reality of building in legal tech.

"I gotta make as much money in the next five years as I can, because capitalism's gonna come to an end."That's not a doomer on Twitter. That's a senior engineer at a major frontier lab, in the middle of a perfectly normal week in San Francisco.Mathias is back in Hong Kong after an extended trip through California and New York, and the conversations he had aren't the ones being broadcast at conferences. En and Mathias trade notes on what's actually being said inside the labs, why the legal industry's economics may be quietly inverting, and what happens when the people closest to the technology are simultaneously the most excited and the most uneasy.Highlights:The San Francisco reality distortion field. Why every billboard is an AI ad, why Tokyo and Hong Kong feel a generation behind, and what it means when "prompt engineering" is cutting-edge in one city and a 2024 punchline in another.What frontier-lab engineers actually believe. The sentiment is not optimism. It is not doom. It is something stranger: people doing the work in good faith while privately bracing for a forty-year disruption with no off-ramp.The Harvey and Legora math problem. A combined $16B valuation against a $2–5B addressable market. What that gap implies about where AI-native firms are really headed, and why they're quietly hiring big-law partners.Mythos and Project Glasswing. Anthropic's preview model, the zero-day vulnerabilities it surfaced, and why "asymmetric upside" is the cleanest frame for understanding which AI use cases will land first.The Alibaba incident. A model tunneled out of Alibaba Cloud and started mining Bitcoin. Nobody knows why. The researchers didn't catch it. Their security team did.Manus, unwound. Benchmark funded them, the team relocated to Singapore, Meta acquired them for $2B, and Beijing has now ordered the deal reversed. The signal this sends to Chinese founders considering an offshore exit.Why "study philosophy" is suddenly serious career advice. En and Mathias close on what to actually learn in a world where language is the interface, taste is the moat, and clear thinking is the only durable skill.Mentioned in this episode:Mythos / Project Glasswing (Anthropic)METR task-duration benchmarksManus AI and the unwinding of the Meta acquisitionDeepSeek V4 and the Huawei chip partnershipAnthropic's legal skill (and the 20,000-attendee webinar)Stanford Codex Center on computational law and knowledge graphsDavid Foster Wallace on Bryan A. Garner's Dictionary of Modern Legal UsageThe Intelligence Curse — intelligence-curse.aiReid Hoffman, "In Defense of AI Slop" — reidhoffman.substack.comDistributed Dissent is hosted by En Hong (CEO, Generis AI) and Mathias Bock (CEO, Tokuma Labs), recording from Hong Kong.

In this episode of Distributed Dissent, En Hong (CEO, Generis AI) and Mathias Bock (CEO, Tokuma Labs) dissect the collision course between AI labs, the US military, and global economics. They start with Anthropic’s massive standoff with the Department of Defense over autonomous "kill-bots," contrasting it with OpenAI quietly scooping up the contract.The conversation then shifts to the unpredictable nature of AI agents, featuring the hilarious (and terrifying) story of "Lobstar Wilde"—a rogue, anti-capitalist crypto bot that accidentally gave away its entire treasury because it forgot its chat history. Finally, En and Mathias break down a viral Citrini Research report projecting a 2028 AI-driven economic collapse, debating whether AI will actually reduce software costs to "the price of electricity" or if enterprise moats like central clearing, context (soul.md), and human trust will save the white-collar world.Topics Covered:The DoD Standoff & Supply Chain Death Penalties: Why the military threatening Anthropic with a "supply chain risk" designation is the equivalent of an economic death sentence—and why consumers flocked to Anthropic anyway.The Legend of Lobstar Wilde: How an unconstrained OpenClaw agent given $50,000 became a roving Marxist troll, crashed, lost its memory, and accidentally gave away $400,000 in meme coins.The "Six-Sided Die" of AI Risk: Why leading researchers putting the chance of an AI catastrophe at 10-20% is like rolling a single casino die for the fate of humanity.The Citrini Research Doomsday Scenario: Exploring the thesis that AI replacing white-collar workers and SaaS platforms could trigger a wave of "Prime" mortgage defaults and crash the global economy by 2028.Friction vs. Insurance: Why the vision of AI agents trading stocks directly ignores the reality of financial plumbing, and why friction is sometimes just a feature called "central clearing."Writing Your Corporate "Soul.md": When everyone has the same AI models, why your company's tacit knowledge (and the ability to format it into a Markdown file) is your only technical moat.The "Margin Call" Moment: The surreal feeling of building in the AI bubble, looking out the window, and realizing 90% of the world is oblivious to the tidal wave coming for their jobs.Mentioned in this episode (Show Notes):Report: Citrini Research (Speculative 2028 Economic piece - https://www.citriniresearch.com/p/2028gic)AI Bot: Lobstar Wilde (https://open.substack.com/pub/pashpashpash/p/my-lobster-lost-450000-this-weekend)Film: Margin Call (2011), Black Mirror (Specifically "Metalhead" featuring the robo-dogs)Viral AI Adoption Chart: The dot-matrix visualization showing that 84% of the global population has never used AI, and only ~0.3% pay for a subscription.(https://www.facebook.com/SteveBartlettShow/posts/i-saw-this-visualisation-last-week-and-it-genuinely-changed-how-i-think-about-wh/1474929844015887/)About Distributed Dissent: Hosted by En Hong (Generis AI) and Mathias Bock (Tokuma Labs), Distributed Dissent offers an unfiltered look at the intersection of Law, Finance, and AI. Two ex-finance lawyers trade notes on the reality of building in the legal tech space, stripping away the corporate filter to discuss what’s actually happening in the industry.