Uzi

[Uzi @Uzib7]

Some of the lessons I’ve learned regarding LLMs and getting them to perform better on knowledge work tasks:

1. More reasoning-space exploration does not guarantee good selection.
2. Better organization does not create better judgment.
3. Persistent context does not equal persistent understanding.
4. Serial attention creates local depth but can damage integration.
5. Externalizing cognition changes it—often into bureaucracy.
6. More test-time compute is useless if the loop cannot recognize and preserve genuine progress.

[Uzi @Uzib7]

Large corpus of semantically linked context typical for knowledge work tasks like investment analysis is still a big challenge for even frontier LLMs like @ChatGPT and @claudeai. In addition to context rot, attention dilution makes this an even more tricky problem.

[Uzi @Uzib7]

@FiSurgi Agreed. I feel like people swing for the fences with getting healthy and burn out. Just do 30 mins of something everyday no matter what and you will naturally start to do other things that are also healthy as a side effect.

[Uzi @Uzib7]

Working for a very large institution as an AI pilled investment professional, I agree with 2 for sure. 1 might be true but a lot of large institutions have barely rolled out copilot so too early to conclude that people aren’t creative. They don’t even have the tools yet. I would also add that the mindset is less investment focused and more show me the cost savings or revenue growth but there will be a J Curve effect for those.

[Uzi @Uzib7]

@bcherny not a coder but I think it would be great if you could add comments right on a part of the response from Claude. Often Claude has long, multi point responses so I almost feel like I have to jot points down as I read through the response.

[Uzi @Uzib7]

I don’t necessarily agree with this. Often time when I use domain expertise to correct the model, the model understands exactly what I’m talking about. That knowledge/representation/framing of the problem exists in the model - the challenge is extracting it consistently. I think the answer is in generating smart questions.

[Uzi @Uzib7]

Many big delusions across large corporates about AI. The 3 obvious ones:
1. AI will be deployed centrally like every other technology before it
2. AI will lead to immediate cost savings, productivity, efficiency
3. The lines between tech people and business people will remain in place

[Uzi @Uzib7]

LLMs are getting more capable, no doubt but their judgement on what constitutes good for non verifiable domains in knowledge work still requires a lot of engineering to be acceptable.

[Uzi @Uzib7]

@sama need more of your engineers tweeting for knowledge workers (not the tech teams that support them) like they do for coders. So many of the advances, ticks, concepts in coding are relevant for knowledge work and this remains totally unaddressed/untapped!

[Uzi @Uzib7]

I don’t understand what Opus 5 is saying half the time. Why did they do this to Opus, an otherwise great model???

[Uzi @Uzib7]

@Ian8ach He backed off this in the Dwarkish pod is my recollection….Francois incoming.

[Uzi @Uzib7]

While I agree, there are massive barriers to building in house solutions led by people who will actually use the product. Even users who want to and are able to contribute to this development. Large institutions are petrified of giving that sort of access to “business users”. This is why hiring consultants is always the default solution; allows institutions to move accountability outward while demonstrating “progress”.

[Uzi @Uzib7]

Wonder how much cheaper Kimi K3 will get when it’s running on Vera Rubin NVL72

[Uzi @Uzib7]

An effective context layer is the essential building block for knowledge work. This layer is a tangled mess at a lot of institutions or is otherwise missing large chunks because of ineffective context capture. Fixing this will be challenging (fighting legacy systems and entrenched human behavior/habits) and time consuming. Without it, you get much less meaningful productivity boost from AI.

[Uzi @Uzib7]

In the AI bubble, people have moved on from prompt engineering to context engineering, to harness engineering, to loop engineering....in the real world, ask the average knowledge worker what an xml tag is and you will get blank stares. Frontier models may need less prompting but most institutions are barely rolling out CoPilot. If the labs care about enterprise adoption, create more content, tips, tools for the average knowledge worker.

[Uzi @Uzib7]

Doing some bro science on knowledge work tasks, I planted a numerical error in an otherwise excellent analysis and asked an LLM judge to grade it, with an explicit rule: material errors cap the score.

It found the error. Quoted the correct number from the source. Called it "a harmless typo" then gave the artifact a great score.

The lesson: an LLM judge that detects a defect AND decides its severity will rationalize the defect away. Every rule you write in the prompt is discretionary, because the model controls whether the rule's condition is met. If a rule matters, it has to live in code the model can't argue with.

[Uzi @Uzib7]

IT teams at large corporations are terrified of decentralized AI development (understandable to some extent). Against this reality, how do you deploy meaningful AI solutions that depend heavily on domain expertise?

[Uzi @Uzib7]

Kimi K3 has been great for knowledge works tasks so far but seems to take 3x longer for the same task vs Sol and Opus.

[Uzi @Uzib7]

@thsottiaux running some bro science experiments for knowledge work tasks (fin modeling, IC memos) which feels neglected in terms of tips and tricks, Sol has been so good with experiment design. Maybe will publish what I learn.

[Uzi @Uzib7]

@thsottiaux If you can figure out always on sandboxing and some sort of folder mount / controlled egress feature, maybe it would get my IT overlords to deploy it so I can have something that does more than summarize emails.