Selman Selmanovic @SSelmanovic96

Has anyone noticed how Anthropic suddenly (after the open-source boom) ships models with much less cost? I mean,i don’t complain,just saying…

Everlaw 2026: 49% of legal pros use gen AI (+12 pp YoY). Easy headline. Harder for Europe/Balkans: when usage jumps, does your stack know which law is in force, or only how to sound confident? Ask vendors for the audit path, not only chat UI.

I’ve been on the Cursor Ultra plan since December 2025 and they built a great product, but thanks to @Da7_Tech I’ve found out about Devin and their model. This is the AI we need.

A model can finish a legal task and still leave something you cannot send to a client. That gap is the product problem. Technically complete ≠ client-ready (jurisdiction, cites, audit path). Eval only on fluency or “task done” ships risk onto the firm’s letterhead.

@HermesWatcher Please put it inside the Hermes when they rollout the API or sub

@harvey Tracked changes in Word: lawyers already live there. Less friction. When an agent edits counsel’s draft, does the authority set stay pinned to in-force text the firm approved, or can fluent rewrite drift the legal basis? UI helps adoption. Version pinning helps sleep.

So I split evaluation: 1) recall: candidate text found? 2) citation-version fidelity: binding version, right article, date and place? Wrong version + high recall = professional-risk error. Relevance alone is the easy half. CG: pogrešna verzija propisa ≠ blaga greška.

Legal RAG demos love recall. “We retrieved the relevant statute.” The failure mode I care about: you retrieved *a* statute, not necessarily the version in force for that jurisdiction on that date.

@gen_legal_inc This framing matters more than another raw accuracy table. “Task finished” and “lawyer can defend it to a client” are different products. Would love to see more evals that fail a run when the cite points to a superseded version, even if the prose looks done.

@everlaw 49% usage is the adoption curve. Curious what share of that usage sits on workflows with an explicit citation / version check versus freeform chat. The report number is useful. The missing operational metric for buyers is still “wrong statute, confidently delivered.”

@street_brawler2 @HermesWatcher Hahahahaha

@thtbee_ Where do you even use Gemini 4 , is it even out ?

@Da7_Tech So,whats your thoughts in comparison of SWE vs Opus,Grok,Fable etc ?

@danliu I find 3.7 better than 3.8

@thdxr The issue is that not everything should be automated by AI.Many people use AI to build software that is inherently AI-native and unnecessary.I’ve experienced this myself; then I realized that some tasks are better to be maintain deterministic processes (ofc built fastly with AI)

If you wrote a book about your life and gave it to someone, would it change their life for the better?

@Mr_Salio Ive been seeing these posts for a while and no anz official announcments.

@Google Legal AI. It is so good with legal stuff. The best model from Google so far. Extremely fast with unbelievably good reasoning.

@pankajkumar_dev Gemini 3.7 flash is the best model so far feom Google. Which issues have you experienced using it?