vas
CEO @varickai | AI Agents for Enterprise: varickagents.com | prev. @meta
Page 1 • Showing 20 tweets
Everyone's building AI to replace employees but is missing something: companies succeed because different employees with different thinking debate and merge into better outcomes. If all your agents run on the same LLM, you've killed that variability. Even across different LLMs, training data overlap is massive. Even after giving each one it's own markdown file with a fake personality, you're getting the illusion of diversity. This will become the cognitive monoculture problem in AI.
"Pretty soon we'll see a 1-person billion dollar business guys!" You will literally never, ever see this. This will never happen. Stop coping. Even if that was possible, it would have to be done by someone smart... and someone smart would just hire help to get to 10 billion dollars, faster. This whole conversation is driven forward by VCs trying to insert themselves in the AI discussion. Same with "10M ARR isnt enough, the new bar is 100M!". Absolute nonsense while trying to sound visionary. Usually followed up with a 'this tweet has been deleted'.
I booked a hotel by speaking with what I suspected was AI It's very impressive to see the AI agent on the other end walk through tool calls - i.e. if I ask for 1 room vs 2 rooms and a room with a view vs bed and breakfast, it knows which data source to look into and how to format that into what is essentially a db query After @polyaivoice dropped their demos yesterday I now realize it definitely was their AI agent If you've been following me for a while you remember when n8n gurus were in full-force last year, explaining how you could set up voice agents for small businesses with n8n templates (lmao). Anyone who tried this knows it didn't work because the underlying logic was too complex for a simple n8n wrap, and the models weren't good enough. You can do this on Poly now. Kudos to the team!
LLMs are stochastic. By extension, Agents should be too. In practice, are they? If I have Codex scan a codebase for bugs, will it uncover the same bugs and miss the same bugs every single time? Temperatures are low for a reason. When you build agents, you’re navigating the tension between determinism (inherently non-agentic, incapable of performing many high-value tasks) and non-determinism (inherently unpredictable, and impossible to trust). Your agent is only as good as you manage this tension. Anyone building in this space?
Product requests - please build this and give me access and I will shill you to my followers: - Text Trainer I give you a ton of my writings (segmented: ie tweets, emails, memos, client deliverables) and you turn that into a masterful prompt/skill that gets as close as possible to replicating that tone and styling, removing any AI-slop. - PDF/Docs to Audio This has been done 1000 times and all of the products suck. I just want to paste the text / upload the file and have it read to me, audio-book style with a human-sounding voice. I also want to be able to have conversation, 'pausing' the 'audio-book' to dig deeper into a specific section or ask question, then 'unpausing'. - AI Repository Houses all of the best libraries/githubs/services for every part of the AI Agent stack (ie voice is X, evals is Y, memory is Z). All in one place, updated constantly as new ones are coming out so I can stay up to date on the latest developments without bothering my CTO and engineers. More to come soon.
AI makes the future fractional Your work will be fractional across a few different organizations, as you delegate to and manage AI systems @naval on a podcast a few years ago said that a gig economy is the future and at the time it wasn't clear what that meant I don't think it is gig work in the current definition of the phrase - where you are contracted by hundreds of different 'employers' over the course of the year (i.e. Fiverr gone mainstream) Instead I think it is fractional, in the sense that you will not be working 8 hours a day on a single function, but rather sporadically throughout the day on several functions (potentially at different companies) For example: I need an A+ Chief of Staff. With AI optimized systems, there simply isn't enough work for them to justify 40 hours a week with me. It then makes sense for them to be a fractional Chief of Staff, working 2-3 hours a day for 2-3 different companies.