Amit Sharma

Building sidekick for lawyers, inside Microsoft Word.

[Amit Sharma @capreal26]

@andrewarruda Or our definition of shitty keeps getting updated frequently?

[Amit Sharma @capreal26]

@abhitwt Human in the loop

[Amit Sharma @capreal26]

Accurate & free (or extremely low cost) visual and audio translation is killer usecase for wearables. Once the players get the form factor right, this will be real personal productivity enhancer. Actually having an AI model speak into your ear, based on what you're seeing - is the real jarvis. Exciting.

[Amit Sharma @capreal26]

@rwalk_xyz 'ChatGTP' at the lowest bar is too real...

[Amit Sharma @capreal26]

@OdedRechavi Reject inferencing?

[Amit Sharma @capreal26]

imo, we should not mind AI written/edited content as long as its solid. Problem is that it regresses to mean v v fast and hard to detect as author. I wish there was a service which could take AI slop and continuously extract signals (if any) from that. If you ask AI to do that, it generates more slop.

[Amit Sharma @capreal26]

@DocPriyamMD Is that the hormone/protein what glp’s act upon?

[Amit Sharma @capreal26]

@scottastevenson @Alfred_Lin Largely agree. Dynamic routing is in the same category as SFT. On point #2, the real value may not be in routing different tasks to different models but identifying the real leverage point task in the workflow, and ensure that output is properly benchmarked/eval'ed.

[Amit Sharma @capreal26]

@SandeepMall Point is that their strategy is well known for quite some time now. Everyone plays to their strengths, and so do they. We need to figure out what our strength is, and play to that.

[Amit Sharma @capreal26]

Great question. I think the answer is somewhere in the middle, as always.

On one hand, the reported ARR is not GAAP revenue and 'annualized' based on most recent month, when Anthropic is doing intense fundraising. Here's a perspective: https://t.co/OjObKjOpGP

On the other hand - there is hardly any tech team in the world which has not bought claude pro ($100/seat/month at minimum) for everyone in their team. Add pro max ($200) and API integrations, the spend on Anthropic's (& OpenAI/Google) AI offerings has skyrocketed - irrespective of who's the IT implementation partner, etc. Teams are seeing massive productivity gains in development work (esp new builds). How that impacts the maintainence / 'Run the business' / BPO type of work is still open because of all the organizational context, politics and (rising) token costs.

Most of the analysts or even leaders continue to parrot Jevons paradox and implementation story (with parallels to cloud) but this is totally a new world and incumbents have absolutely no clue or fresh narratives. They keep peddling how IT industry has survived many innovations, etc. etc but none of them are doing meaningful work close up. Have you seen a solid case study / demo / product from a single large IT co, in last 24 months, which demonstrates how they have leveraged AI to deliver their work better than earlier. Its all slideware and fluff, right now.

[Amit Sharma @capreal26]

@willchen500 Insurance company using Roy Lee’s playbook. Phew!

[Amit Sharma @capreal26]

Makes complete sense for the top firms to build their own, on-prem AI capabilities. VC funded Legal AI firms are rolling in valuations on the back of two things:
1. Marquee client logos signalling credibility
2. Powerful AI capabilities & harnesses (custom built for bespoke legal work).
#2 is now less & less of a moat, and an aggressive law firm can assemble a great team of technologists & build it out. Many 'Forward Deployed Engineers & Legal-Technologists' are anyways from these top tier firms & practices only. Keeping your client's data off 3rd party servers (and protecting privilege) is the biggest benefit.

This will be a reality in most domains, not just law. Top firms will lead the way, and many will follow. There will be blow-ups and disasters but Anthropic/OpenAI/Google have ensured that the end customers have much more leverage than the rest of the value chain. Old adage of: 'Will incumbents get innovation before start-ups get distribution' is passe. All incumbents now have access to innovation, and some lead it as well. Real question now is: Will the incumbents still be excited and committed about AI when its no longer sexy and foundational model labs are not dropping new models every month. It will happen and 2025-26 is definitely an aberration. Incumbents who can shed their deadweight & really pivot to a culture of experimentation and fast deployment using AI will last out way beyond this unusual period.

[Amit Sharma @capreal26]

@mtrajan Prediction that model release frequency goes down significantly post-IPO?

[Amit Sharma @capreal26]

@mtrajan Frankly, AI Deployment Specialist would make more sense but then it wouldn't behoove a CEO's post.

[Amit Sharma @capreal26]

Great advice for cs and math undergrads feeling overwhelmed by ai

[Jimmy Heaters @CathPoaster]

new grads often ask me what they should be doing so they don't fall behind in the ai space. there's a lot, but its honestly super manageable. become intimate with model internals. proof based linear algebra. non-convex optimization. this is stuff you could've done in undergrad. it definitely takes some time and work, but its doable. have taste, have opinions. train a small model, then train a big one. vLLM internals, tensor parallelism. hand roll kernels. cluster orchestration. do you have opinions on synthetic data? why don't you? SFT, PPO, you should know this. learn Triton. everyone is reproducing papers now so you need to be doing more. do you know the semi supply chain? where are the bottlenecks? hardware, man, hardware. your little gpu rig erector set in your basement isnt gonna cut it. build a cluster, a big one. pretrain a 800B model. now postrain it. serve it to millions of people. you should be able to beat deepseek on some benchmarks now. its a lot to take in but it all snowballs. this what job security looks like from now on. do you want to work in tech or not

[Amit Sharma @capreal26]

@ZachAbramowitz Well said. Simplification is a worthy aim.

[Amit Sharma @capreal26]

@MorganBarrettX Do they have the IP rights to client's data/docs? If they use it to train AI, they lose privilege (US v Heppner)

[Amit Sharma @capreal26]

@SellersCounsel Absolutely. People routinely underestimate how different the client management practice/business is from running an AI business.

[Amit Sharma @capreal26]

Makes complete sense for the top firms to build their own, on-prem AI capabilities. VC funded Legal AI firms are rolling in valuations on the back of two things: 1. Marquee client logos signalling credibility 2. Powerful AI capabilities & harnesses (custom built for bespoke legal work)

#2 is now less & less of a moat, and an aggressive law firm can assemble a great team of technologists & build it out.

Many 'Forward Deployed Engineers & Legal-Technologists' are anyways from these top tier firms & practices only. Keeping your client's data off 3rd party servers (and protecting privilege) is the biggest benefit.

This will be a reality in most domains, not just law. Top firms will lead the way, and many will follow. There will be blow-ups and disasters but Anthropic/OpenAI/Google have ensured that the end customers have much more leverage than the rest of the value chain.

Old adage of: 'Will incumbents get innovation before start-ups get distribution' is passe. All incumbents now have access to innovation, and some lead it as well.

Real question now is: Will the incumbents still be excited and committed about AI when its no longer sexy and foundational model labs are not dropping new models every month. It will happen and 2025-26 is definitely an aberration. Incumbents who can shed their deadweight & really pivot to a culture of experimentation and fast deployment using AI will last out way beyond this unusual period.

[Ira Rothken @rothken]

Kirkland with a 500M investment in AI legal is important in signaling a big advantage and controlling their own AI intelligence layer and hybrid search and document stores.

Harvey and Legora, while very robust and helpful early wrapper platforms for firms needing a quick AI infusion, may end up being an AI concept prover for some firms -but still hugely helpful to firms without the budget and talent to build their own platform. Once big firms ramp up they may very well see that they need to own their intelligence layer and workflows and market that they are different and better or more AI native than other firms in the AI tech space who use only mass AI legal vendors.