Robert Käll (@robertkall) header
Robert Käll (@robertkall) profile picture

understand the world using data ### MEMES ABOUT BUILDING AI APPS AND ENTREPRENEURSHIP EVERY TUES & THURS ### ❤️👨‍👩‍👧‍👦🏠🎼✈️🚀📈🎸⛷️🥊😃🌶️🐿️🍺❤️

Intercontinental
Joined June 2023
Photo by Robert Käll (@robertkall)Photo by Robert Käll (@robertkall)Photo by Robert Käll (@robertkall)Photo by Robert Käll (@robertkall)Photo by Robert Käll (@robertkall)Photo by Robert Käll (@robertkall)Photo by Robert Käll (@robertkall)Photo by Robert Käll (@robertkall)Photo by Robert Käll (@robertkall)Photo by Robert Käll (@robertkall)
I've built companies long enough to know that mistakes are unavoidable. And I genuinely don't mind people making mistakes while moving quickly and trying new things. What frustrates me is when the same mistake happens repeatedly because nobody captured what went wrong, changed the process, or verified the fix. A mistake should produce learning. Learning should change behavior. And changed behavior should improve the outcome. Otherwise we're just paying tuition for the same lesson over and over again. #Understanding #RobertKall #CienAI
1
6
When you build an agent that can use tools, call APIs, execute workflows, and retry failures, you're giving it genuine autonomy. That's powerful, but it introduces a different class of engineering problems. An agent can get stuck repeating an operation, consume enormous resources, or continue pursuing a task that should have been stopped. Production systems need budgets, execution limits, timeouts, approval gates, and the ability to interrupt or recover safely. The goal isn't to make agents less capable. It's to make their autonomy predictable enough that customers can trust them. #GuardRails #CloudCost #RobertKall #CienAI
1
8
One of my biggest frustrations as a founder is watching talented people jump straight into execution without understanding the problem. Before acting, I want three questions answered: What has been happening? What is likely to happen? What would happen if we changed something? Those questions aren't bureaucracy. They're how you avoid spending days solving the wrong problem. I would much rather someone tell me they don't understand something than confidently deliver a solution to a problem we never had. #ActivityVsOutcomes #RobertKall #CienAI
1
4
34
One lesson I've learned from working extensively with AI agents is that conversations are a terrible long-term knowledge-management system. You explain a problem, make decisions, establish conventions, and eventually all that valuable context becomes difficult to retrieve or reuse. We increasingly move important knowledge into structured Markdown documents, skills files, and reusable knowledge sources. That allows agents to start with the right context instead of rediscovering the same lessons. The conversation is temporary. The knowledge created during it should become a permanent asset. #Skills #Memory #RobertKall #CienAI
1
4
45
"Seems a little overengineered." is my favorite prompt to Claude/CODEX when I ask for an idea how to solve a problem. It almost always agrees and provides a simpler and more reasonable solution.
2
96
I've worked with plenty of smart people who have learned to equate being busy with being productive. Meetings, reports, initiatives, and completed tasks can create the appearance of progress. But customers don't pay for our activity. They pay for outcomes. As a founder, I want to know what changed because of the work. Did we improve something measurable? Did we learn something that changes our next decision? If not, we may have been very busy accomplishing nothing.
1
9
Robert Käll retweeted
CNN projecting that SQL will still be the data language of the future
10
18
3
172
7,646
Are you mfs still going to call it SI? Yeah, I didn't think so...
1
8
Robert Käll retweeted
Founders: Executive demo strategy: - Max 30 minutes - 3 slides upfront - 1 concrete ROI example - 7 minute product tour - Clear next steps CEOs don't want features. They want: - Revenue impact - Cost savings - Strategic advantage Keep it high level.
12
10
2
179
9,517
I've become increasingly convinced that choosing the smartest model isn't necessarily the most important decision in an agentic application. What can the agent actually access and do? Can it retrieve the right data, query systems, execute code, validate results, and take action? This is why tool integration and protocols like MCP are so important. A slightly less capable model with excellent tools and context can be far more useful than a brilliant model trapped inside a chat window. Intelligence without the ability to act has limited business value. #Agentic #Tools #RobertKall #CienAI
1
13
I am a real engineer :)
You're not a real engineer if you don't have a mystery $0.53 AWS bill coming from a rogue service in perpetuity.
9
Robert Käll retweeted
You're not a real engineer if you don't have a mystery $0.53 AWS bill coming from a rogue service in perpetuity.
396
123
104
4,176
176,487
So I have posted this graph twice before in the last 12 months. But the jump and being way over my projection is interesting (I am at 60% and last year I projected 35%). I am not interacting w/ AI during calls and most meetings, but the rest of my workday is almost all AI agents. The main reason is that I have shifted towards doing a lot of (AI-assisted) coding again. But I am also using it in almost all document and plan creation, plus more and more other biz workflow automation. #AIAdoption #Agents #CienAI #robertkall
1
1
22
update - it's now 80%
1
this makes no sense...
LLMs in space! @Google and @planet put TPUs on a satellite with project Suncatcher, and now we’re running AI inference from orbit ✨
19
Robert Käll retweeted
LLMs in space! @Google and @planet put TPUs on a satellite with project Suncatcher, and now we’re running AI inference from orbit ✨
A ‘hello world’ milestone for Project Suncatcher 🚀 After launching our prototype satellite a few weeks ago, we’ve now successfully sent a command up to orbit, booted up the TPU, ran the prompt, and received our first complete response back on earth - a haiku!
2
2
1
16
680
One of the easiest traps when building technology companies is falling in love with what you're building. Engineers love interesting problems. Product teams love features. Founders love their vision. But customers aren't buying our roadmap. They're buying an outcome. At Cien, we've increasingly focused on whether customers actually realize the value our products identify. That means measuring what happens after deployment, learning where execution breaks down, and changing the product accordingly. A feature isn't successful because it shipped. It's successful because it helped someone achieve something valuable. #RoadMaps #Outcomes #RobertKall #CienAI
1
1
18
One of the biggest challenges building agentic applications is that an LLM can produce an incredibly convincing answer without knowing whether it's correct. Asking the same model whether its answer looks good isn't sufficient. You need external checks: deterministic calculations, reference data, business rules, test cases, and sometimes independent model reviews. For our applications, the definition of success matters as much as the task itself. An agent that produces an answer is interesting. An agent that can demonstrate that its answer satisfies the requirements is useful. #AgenticQA #RobertKall #CienAI
1
1
17
For most of my career, scaling a company meant delegation. You hired specialists, divided responsibilities, and built processes to coordinate their work. I'm increasingly doing the opposite. With AI, I can take a problem from understanding through design, implementation, testing, and delivery without handing it off five times. I'm finding that this is often faster and produces a more coherent result because the context stays with one person. I'm also looking for individual contributors and leaders who can operate this way. Collaboration still matters, particularly for complex decisions and independent review. But the ability to own and solve a problem end-to-end is becoming dramatically more valuable than the ability to coordinate people who each own a small piece. #Coordination #Handoffs #RobertKall #CienAI
1
23
When building AI applications, it's tempting to let the LLM do everything. But why ask a probabilistic model to do something ordinary code can execute reliably every time? We increasingly separate responsibilities. Let the LLM interpret ambiguous information, reason, and handle novel situations. Let deterministic code enforce rules, perform calculations, validate outputs, and execute repeatable operations. The best architecture isn't necessarily the one with the most AI. It's the one that uses intelligence where intelligence is needed and predictable code everywhere else. #Agentic #Deterministic #RobertKall #CienAI
1
17