
Founder @litvue building real autonomous legal work product
Joined September 2017
- Tweets8.2K
- Following733
- Followers957
- Likes18.9K
Safety and alignment are the brakes on the AI race car
You think they’re there to slow you down but it’s really the opposite.
OpenAI Is Facing a Serious Internal Crisis
They're losing a lot of talented people Nearly 5 people have left in just 14 hours on top of several others who departed over the past week
It makes sense Just two days ago OpenAI published a major safety update saying it's slowing down model releases, pausing internal training when necessary and even holding back it's models until they meet its safety standards
Now we're seeing people leave across different teams including safety alignment research and product Idk if these things are connected but the timing raises some serious questions
Something much bigger is happening behind the scenes at OpenAI internally looks like team are facing a lot of pressure
Been thinking about this since yesterday and @giyu_codes is right.
You need to completely reframe how you think about ai enablement. Its use in an enterprise or mid market org needs to be as ubiquitous as use of the internet.
AI is not something that works when it’s bolted on workflows or even when it automates an existing workflow. It upends the way that things have always been done.
Thats scary, but when you see what’s possible you will never go back. I promise.
Replying to @sethtjf
Full AI enablement for a 500-1000 mid market/enterprise org. Every inch of the company needs to be in the scope of the work, which an agency should be able to communicate to me on how they would do that.
Legit everyone was saying "Well we should start with one workflow, break it down, meet with the stakeholders and review the ROI".
Which yes it's a good fair answer, but they just showed me they aren't ready for a bigger scope then.
141
Incredibly fed up with @Rippling and ready to move off. What's worth moving to?
439
> As a result, standards are inherently subjective, which favors bespoke solutions tailored to specific customers over one-size-fits-all products.
@solirvine absolutely nails it
My takeaways from this:
- Ground truth in legal is elusive because human lawyers are inconsistent in both performance and grading.
- As a result, standards are inherently subjective, which favors bespoke solutions tailored to specific customers over one-size-fits-all products.
4
6
1,444
Replying to @andrewarruda
The cool thing about this being a model config is they can all use it! I’m really just trying to keep repping the YC S15 legal tech crew 😂
1
619
Is this not a public display of ToS violation? lmao
The leading observability platform wants $83,700/yr just to monitor your app and catch issues.
So Replit Missions built DataCat: github.com/replit/datacat
100% free.
This isn't a prototype. Agent swarms explored thousands of workflows in the leading observability platform and rebuilt its features 1:1.
Co-sign
This is the AI deployment hill I will die on:
Subject matter experts should not own the product.
My company Atherial has done over 60 AI deployments. I learned this the hard way.
Subject matter experts should give the playbooks and know-how via discovery interviews, creating a rubric, and writing evals on sets of outputs.
They should not design the solution or write the requirements.
Leave the product roadmap to the SME and you get the obvious automation (copying and pasting between tools) and intern-level AI. Worth doing but not where the value is.
Here's our approach now.
You own:
• Business strategy
• Industry knowledge
• Functional expertise
Atherial owns:
• AI strategy: finding high value use cases
• Solution design and product roadmap
• Engineering
You want an FDE or AI consultant who understands why strategy, product, and engineering all matter, and who owns what.
Product management is a real skill. AI application companies pay a lot of money for it.
The best product people are mini founders. They find big and valuable problems and design solutions that take full advantage of what AI can do.
Defining what good output looks like is step one because software/harness you build changes as you test and hill climb.
Your experts should help train the agent and judge its work. FDEs who know product and strategy should decide what to build.
4
2,955
I spent 4-6 months of this year exploring this concept deeply only to conclude it was a giant waste of time. Maybe I should revisit.
This quoted post is unavailable.
Finally caught up with @abraham_jmb live today and it was 1000% better than I expected (and my expectations were highhh)
2
4
343
Spend enough time close to legal work and ai adoption and you realize that all of this is deeply intimate to the firms and lawyers and no solution (read, vendor products) is truly one size fits all.
There's a lot of vague posting in legal tech right now. "We'll automate this." "This does it end to end, perfectly, every time." The reality is a lot more nuanced, and you see very few real examples.
Estate planning for example is easier to automate. They were already templates. The context per engagement is limited and the rules are basically if then statements.
Most of law isn't like that. When the answer is "it depends," the model needs a lot more context to fill in the blanks about what it depends on and models getting smarter is partly a detriment here.
They now know when a dependency sits outside the context they were given, so instead of guessing, the output stalls or goes off track.
That's why the big legal tech companies are sitting on older models. One of the biggest was defending still using Gemini 2.5 pro a few months ago. It isn't cost. They tuned a consistent result with the harness and the context they built at the time, and every new frontier model means redoing all of that harness engineering.
My prediction is a plateau in publicly released models. Anthropic keeps giving the frontier away. Kimi K3.1 Pro is distilled from Opus 5.5. It's the soul of it, doing Opus 5.5 quirks.
Kimi is an amazing model. The cost per unit of intelligence value is phenomenal.
So we hit a fork. Do we want more intelligence, or a model that's better at the task and context we're actually asking it to handle?
We'll pay more for the second one, even while the cost per unit of intelligence per token plummets.
What we want is predictable intelligence: a system that doesn't change every three weeks. In legal, you can't redo your harness engineering every three weeks. And it's not unique to legal. It's all of us pushing the frontier day after day looking for consistency.
#legaltech
For knowledge work we’re at the stage where your initial question should be asking how a human would do this task and then work backwards to apply an agent to it.
This works especially well greenfield tasks.
4
8
507










