Models are solved, context is the frontier

I just got back from two weeks in the US. San Francisco for HumanX, one of the biggest AI conferences in the world. Atlanta for private equity client work. Landed home Sunday, and the jet lag is real.

Here's what I'm sitting with now that I'm back.

I walked the floor at HumanX and kept seeing the same thing. Data analysis. SQL. Postgres. Finance. Accounting. Admin. Booth after booth of startups building thin features on top of a model.

I'll be straight up. Most of those startups are there to build a little something, raise a round, and exit fast. Because the moment Claude or OpenAI ships the same feature natively, that startup is gone.

That is exactly what keeps happening.

Why would I work with a small startup when I can use Claude Console directly? Or the new Managed Agents feature from Anthropic to host my agents?

If you run a mid-sized company with 50+ people, you should not be betting on an early stage startup to solve a core problem. The big platforms are shipping too fast.

One exhibitor stood out: Vercel.

Vercel had a proper presence at HumanX. A booth. A partner showcase area with a lineup of companies they work with. A workshop with a panel and a talk. Of everything on the floor, the Vercel zone was the most interesting thing I saw.

And Vercel is serious. I'm already hosting our website, our prospect pages, and our Agent Gateway on Vercel. I stream agentic conversations into Slack with the Vercel Chat SDK.

Now I'm looking at Vercel Workflows.

Real question I'm sitting with: do I still need trigger.dev if Vercel Workflows covers the same ground? We're already deep in the Vercel ecosystem. Fewer tools means less maintenance, less cognitive load, less to learn.

Another thing worth noting. Vercel AI SDK is model-agnostic. So I'm not locked into a single model family.

Two updates from the trip:

More on both soon.

While in SF, I visited the OpenAI offices. Shout out to Larry from Startup Partnerships for the hospitality, and to Tyler Scriven, our advisor, for making the meeting happen.

The office is architecturally wild. A building floating inside a building. Brunch with chefs, baristas, every type of food. Very cool space to work in.

We talked partnerships and future collaborations. There is something very interesting forming there.

On my first Friday morning in San Francisco, I woke up to the news that Sam Altman's house got attacked.

A 20-year-old attacking someone over anti-AI politics is sad. It's like attacking an electricity company because you do not want electricity.

AI is here to stay. Attacks and protests are not going to slow it down.

Are there valid concerns about speed of adoption, and regulation?

Of course.

But here is the pattern I keep seeing:

I'm in the third group. You are too, otherwise you would not be reading this.

And this is my manifesto: I'm not saying AI is better than you. I'm not saying you are better than AI.

I'm saying you plus AI are better than either one alone.

After SF, I flew to Atlanta for the first time. Nobu Hotel. Beautiful city. I have private equity clients there where I act as the fractional Chief AI Officer.

The work is two things now: behavior change programs (what we used to call training) and building agentic systems for portfolio companies.

I'm running a cohort for 26 private equity firms. 54 executives in total. One of them was skeptical. Smart guy, great energy, but deeply against AI in the beginning.

"Shit in, shit out," he kept saying. "Whatever I ask, the answer is shit. I don't trust it."

So I asked him one question. What are you giving it access to? Nothing, he said. I don't trust it yet.

I told him: you want a great answer from a tool you give no information to. What exactly are you expecting?

Then I showed him my own setup. My deal pipeline in Notion. Claude with connectors. How I draft follow-ups, summarize meetings, find patterns across deals, write proposals, all grounded in my real data.

He is a full convert now.

And this little story is the whole point of this newsletter.

The models are no longer the bottleneck. Opus 4.7 is incredible. GPT 5.4 Pro is incredible. Tool calling used to be the weak spot. Now it's very, very good.

If you ask an AI a question with no context, it answers from its training data. If you give it web search, it adds live information. For most people, that's it. That's the entire relationship.

But most businesses are sitting on context and do not realize it.

PDFs. Agreements. Spreadsheets. Emails. Image-heavy docs. Meeting transcripts. Support tickets. Resolved customer complaints. Sales emails. Old proposals. Diligence memos.

The context is everywhere. And because it's everywhere, it's also nowhere.

Getting insight out of one PDF today usually means:

Now imagine this instead:

That is where the magic lives.

The context you need depends on who you are:

Andrej Karpathy has been talking about his Obsidian vault connected to Claude Code. Obsidian notes interlink, which gives the model a richer graph to reason over.

I'm asking myself the same question right now. Do I stay with local markdown files synced to Notion, which is our business source of truth? Or do I build a separate Obsidian vault for founder-level personal context, linked back to Notion?

I don't have the answer yet. I'll tell you next week.

Two stories.

Personal finance. I've been bad at personal finance forever. Always tried, always slipped. Last week I dropped my bank statements and last year's tax return into Claude Code and said: help me understand where my money goes and where I can cut.

It found subscriptions I had forgotten about. Flagged recurring spend I did not need. Broke down the kids' school costs. Told me where my money actually goes in a way no app I have ever paid for has.

The 15-minute MVP. A friend of mine is not a developer. He wanted to build something. He started a session with Claude Code. Claude gave him a four-week project plan and he got discouraged.

I told him: remind Claude that it's not a human. It does not need human time. Go back and tell it you have 15 minutes.

He did. Claude shipped an MVP in 15 minutes.

Not the final product. Not polished. But a working MVP. I'd bet that with four hours of real work, he'd have a real product.

The building is not the hard part anymore. The limit is your imagination.

I've been quiet. No newsletter for two weeks. Slow on YouTube. Slow on LinkedIn. It's not laziness. It's the opposite.

Since landing Sunday, in about one and a half days, I have cleared work that used to take weeks. Client follow-ups. Partnership conversations. Agreements. Contracts. Proposals. Custom landing pages for those proposals. All of it.

I remember what "post-conference week" used to feel like with a sales team. Hundreds of business cards. Days of data entry. Weeks of follow-up. That world is gone for me.

Real ask for you this week.

Reply to this email. I read every one. Ask me questions. This community is the reason I keep showing up.

Reply with what you want to see next.

Ciao,

Tim