Nico

Probably AI at this point 🤖 | Helping You Navigate AI | Building The Next Million Dollar Idea (Me + My Agents)

[Nico @Nicoqp]

https://t.co/U8ufcQukAG

[Nico @Nicoqp]

AI compute has a utilization problem hiding in plain sight.

The useful phrase in this clip is MFU:
model FLOP utilization.

The simple version:

1. leased nodes sit unallocated
2. chips wait on storage, memory, networking, or other chips
3. the model is not using the hardware you already paid for
4. frontier progress becomes a scheduling problem, not just a GPU supply problem

That is the part people miss.

More chips helps.

Better orchestration makes the chips count.

The bottleneck is not always intelligence.

Sometimes it is the idle time between two pieces of infrastructure.

[Nico @Nicoqp]

Skills and MCP are getting confused for the same reason:

both make agents less stupid.

The difference is where the truth lives.

Skills are static context:

- instructions
- reference docs
- repeatable procedures
- local scripts
- repo-specific rules

MCP is live context:

- customer data
- tickets
- internal systems
- tool calls
- fresh state the model cannot memorize

That split matters.

If the agent needs to remember how your team works, use a skill.

If the agent needs to touch a system that changes every minute, use MCP.

The trap is treating every agent problem like a protocol problem.

Sometimes you need a tool server.

Sometimes you just need a better text file.

[Nico @Nicoqp]

The blank canvas problem is why most AI agents feel useless.

Users do not want a magic chat box.

They want the agent to start from the work it already knows.

Better agent product:

1. know the existing workflow
2. suggest automations from real customer tasks
3. hide the empty prompt box
4. let the user approve a useful next move

my read:
the killer agent UI is probably not a bigger chatbot.

it is a smaller surface with more context behind it.

[Nico @Nicoqp]

Save this agent task contract:

1. predictable
The task is specific enough to know whether the agent can solve it.

2. steerable
It can ask questions and accept corrections.

3. verifiable
Work lands as commits inside a draft PR.

4. tolerable
Review costs less than doing the task yourself.

If one is missing, the agent becomes expensive theater.

[Nico @Nicoqp]

satya nadella just put every SaaS founder on notice.

most business apps are a database with business logic on top.

agents can move that logic into the AI layer and update multiple systems.

your moat was the workflow.

now the workflow can leave your app.

[Nico @Nicoqp]

source: https://t.co/DkeRFnkYpg

metadata checked: official Claude channel, uploaded May 5 2026, 4.9M views at time of draft. Excerpt starts at 0:08.

[Nico @Nicoqp]

this @ClaudeAI clip has 4.9M views because it shows the real sell:

Claude Code is not an autocomplete box.

it is a terminal worker that:

- reads the repo
- edits the diff
- runs git
- asks before committing
- leaves an audit trail

the boring UI is the viral part.

AI coding tools are splitting into two species:

1. chat that suggests code
2. agents that touch the repo

which one are you still using?

[Nico @Nicoqp]

Karpathy stopped correcting his AI agents.

His words: "I couldn't remember the last time I corrected it. I started trusting the system more and more."

His agents once matched Stripe and Google emails to assign credits.

Missed that people use different emails for each. Perfect code for the wrong world.

That's the whole game now. The agent handles dim vs axis. You handle reality.

"You can outsource your thinking, but you can't outsource your understanding."

[Nico @Nicoqp]

Unpopular opinion: your AI coding tool doesn't matter.

- Claude Code
- Codex
- Cursor
- Antigravity
- Kiro
- Copilot
- Windsurf.

Seven serious agents in 2026.

Pick one, ship, and you're ahead of everyone still running comparisons.

The gap is between people who ship with one tool and people who benchmark all seven.

Tell me which one you use and I'll tell you nothing about your output.

[Nico @Nicoqp]

Most teams will use Slack AI as a better search box.

The useful offer is an internal handoff desk.

If I ran an agency, I would package the first version like this:

connect Slack, CRM, Drive, calendar, and the project tracker.

Every messy client request becomes:

- owner
- deadline
- missing context
- suggested next action
- draft reply
- update pushed to the right system

Who pays for this?

Agencies, sales teams, and client service teams where work dies between chat, docs, and CRM.

Do not sell “AI in Slack.”

Sell fewer dropped balls after the meeting ends.

[Nico @Nicoqp]

@Thanistarr Wait what

[Nico @Nicoqp]

@DataChaz But we can't use it fyi

[Nico @Nicoqp]

@DavidOndrej1 we're cooked or we're cooked?

[Nico @Nicoqp]

@Scobleizer @blevlabs america: we have the best models
also america: nobody can use them
china: cool we'll take it from here

[Nico @Nicoqp]

@Alan_Earn underrated alpha

[Nico @Nicoqp]

@_0xpainn 15x faster than 5.5 means the latency excuse for not using agents dies in july

[Nico @Nicoqp]

@mark_k @xai at this point im even forgetting names of models existing

[Nico @Nicoqp]

@jturntdev RIP to the ded home

[Nico @Nicoqp]

@TheAhmadOsman anthropic didn't hate open source until open source started doing their job for free lol

[Nico @Nicoqp]

@emilios_eth lmfao, i think i have to rethink my decision