Daily AI Fresh News · Nov 3 Evening Edition: Pando Rethinks Git & KVCache Buzz

Lots of new stuff in AI today — from a European fund to Apple Watch, and OpenAI quietly building infrastructure.

Today we'll cover a few things: first, a new European fund; second, Claude Code scanning repos; third, a restaurant vision system; fourth, a Cursor quickstart; fifth, Apple Watch AI.

Then in the deep dive we'll talk about three things: the KVCache meetup, the new Ollama release, and OpenAI's Pando.

Yeah, I'm pretty keen to talk about all of those.

Alright, let's run through the news quickly first, then dig into the big ones.

First up, I just saw on TechCrunch that Yvonne Bajela launched Evertrue Capital with a $17 million first fund.

Investing in early-stage AI and fintech — the European startup scene is clearly still hot.

$17 million isn't huge for an early-stage fund, but the timing is pretty interesting.

European AI funding has been pretty conservative these past two years. Someone willing to jump in now means she sees a window.

And she's specifically targeting AI plus fintech — those are the two easiest areas to land and scale profit, right?

Second, there's a video on YouTube about using Altimate MCP to find idle Snowflake warehouses in Claude Code.

Basically letting a coding agent check which data warehouses are just sitting there burning money.

I'm pretty bullish on this direction. The point isn't the checking — it's baking cost governance into the agent's workflow.

Think about it: idle warehouses are pure waste. Have an agent scan regularly, and whatever you save is straight profit.

Using MCP as the interface so Claude Code can connect directly to data platforms — this kind of thing is really popping up lately.

Third, there's a very down-to-earth post on Reddit — someone wants to build a production-grade vision system for restaurants.

Two things: calculate seat occupancy, and tell whether a table is clean. They're asking for architecture advice.

Seat occupancy is manageable, but judging whether a table is clean — the labeling criteria are a huge trap.

Right, what counts as clean? Does one french fry count? That's a product definition problem.

So the real challenge in these projects is never the model — it's translating business standards into labelable rules.

Fourth is lighter — a 60-second video on YouTube called 'Get started with Cursor AI in five steps.'

One minute to learn Cursor — the pace is really fast.

The fact that this kind of content takes off shows the tool has gone mainstream — the audience isn't just programmers anymore.

Makes sense. Shorter tutorials are actually a signal that the user base is expanding.

Fifth, also TechCrunch — Apple Watch AI features are entering testing.

It's called Audio Intelligence — you can rewind conversations and replay them, plus get summaries and take notes.

That's actually a real need. The watch is close to your ear, so it's a natural fit for conversation memory.

But doesn't it feel a bit dicey? It's recording all the time — how do you address privacy?

The key is how much processing happens on-device. If it all goes to the cloud, controversy is inevitable.

Alright, that's the news. Now let's dig into a few more technical topics.

First deep dive: KVCache Meetup Korea 2026, held September 30 in Korea — the recap just came out.

Organized by PyTorchKR and a startup called XCENA.

Here's a key point: why does KVCache deserve its own meetup? Because it's the biggest chunk of inference cost.

The longer your context, the more this cache balloons — it eats up GPU memory. So optimizations around it are literally money.

Wait, so when a startup like XCENA jumps in, are they selling hardware or solutions?

Given the vibe of a community event, probably storage or memory hierarchy — finding cheaper places for the cache to live.

So it's about giving KV cache a more cost-effective home instead of always piling on HBM.

Exactly. Tiered storage is going to become more and more mainstream over the next year or two.

Second, Ollama released v0.40.14 at 1:40 AM — the version numbers are jumping fast.

Just pushed on GitHub, no big fanfare.

That's Ollama's rhythm now — small, quick releases, focused on polishing stability and compatibility.

Don't dismiss it as a minor version. In local inference, whoever saves users from headaches keeps them.

So its moat is actually ease of use, not performance specs.

Exactly — the deployment experience is its product strength.

Third, the big one: OpenAI built Pando, a Git backend made specifically for their giant repos.

They say median push latency dropped by nearly 80%, and now they handle close to a petabyte of Git data daily.

This is underrated. Everyone's watching the models, but engineering infrastructure is their hidden moat.

Pando's core is processing branches independently and reusing existing packfiles across pulls.

Reusing packfiles — I get that. Basically don't re-send the whole pack every time.

Right, and it's Git-compatible — not a total rebuild. That's smart.

But I wonder — is something at this scale only feasible for a company the size of OpenAI?

You can copy the ideas, not the scale. But the independent branch processing — mid-sized teams can absolutely borrow that.

So the takeaway: an AI company's competitiveness is half model, half invisible engineering foundation.

Completely agree. Models get caught up to; infrastructure leads last longer.

Alright, that's it for today — plenty to chew on.

Yeah, see you tomorrow.

See you tomorrow!