Serge Jespers — Journal

Are we in the era of "good enough"?

Sat, 24 Oct 2026 12:00:00 GMT

The iPhone Duo is heavy. It dropped Face ID. No telephoto. It gets warm.

And here’s the part that gets me: Xiaomi, although banned in the US, launched a foldable two days earlier in basically the exact same shape. It has a periscope telephoto. Better selfie camera. It’s lighter. It’s cheaper.

So “you can’t fit that in a foldable” isn’t the reason. Apple decided what made the cut, shipped it, and people lined up around the block anyway.

I don’t think that’s laziness. I think it’s the clearest example yet of “good enough” as a strategy. When people trust you, they buy the direction, not the spec sheet.

I often live in this world. Event dates don’t move. When doors open at 9am, the activation is either running or it isn’t. So at some point you stop asking “is it perfect?” and start asking “what’s the one thing that has to work when someone walks up to it?” That gets all the polish. A lot of times, the rest gets “it’ll hold.”

The difference is Apple has decades of goodwill to spend when v1 is rough. Most of us don’t. Ship something half baked without that cushion and nobody sticks around for v2.

Where do you draw that line in your work?

Today's talk of the town is a model that can't talk. A note about Jev.

Sat, 19 Sep 2026 12:00:00 GMT

Beside my day job, I run a one-person operation with an AI staff: agents that triage my inbox, track trends in melodic house, watch my servers, and brief me every morning.

Here's the part nobody tells you about running agents at this scale: most of what they do isn't reasoning, it's tiny decisions, made thousands of times a day. Which folder does this email go in? Does anything need me right now? Is this "news" actually new? Pointing a big language model at every one of those questions is slow, expensive, and occasionally confidently wrong at 2am.

So I stopped asking an LLM to reason through everything. In comes Jev, a model TypeSafe AI dropped just a few days ago, still in early access, built for exactly this kind of grunt work.

Jev is what they call a "System One Model." Instead of generating text token by token, you define the possible answers up front, it scores all of them in parallel, and hands back a calibrated probability for each in under half a second. Fixed output space means it can't hallucinate and can't send back something malformed. It's also absurdly cheap: $0.042 per million input tokens, output free. A frontier LLM runs $0.20 to $10 per million, plus output cost on top. For something polling every 30 minutes forever, that's the difference between a tool you keep running and one you quietly kill after a week.

So now, instead of "figure out what to do with this email," I ask atomic questions: "Which mailbox?" "Needs a reply?" "Urgent?" Each comes back as a number. My code decides what to do with the number.

Three places it's running right now:
Mail triage: Jev shadows my rules engine daily, and I track agreement over time. It has to earn its way into actually touching my mailbox.
A trend gate: my music trend watcher polls every 30 minutes. Jev decides if anything material changed. No change, no noise. It fails open, so it never misses real news.
An urgency watchdog: scans my inbox every 30 minutes, only speaks up when it's actually confident. Silence is the win condition.

None of this is about a smarter model. It's about the shape of the decision: judgment as a cheap, metered API call instead of a chat window. Save the big models for the interesting problems, give everything routine a number and a threshold, and let code do the rest.

More info here: https://typesafe.ai/

We’re not sleeping on Chinese open-weight models. We’re just not hearing about them.\ \

Sun, 30 Aug 2026 12:00:00 GMT

fal.ai just shipped H3 Max. It’s the #1 ranked AI video model right now. Beats Veo, beats Kling, beats Sora on quality and prompt understanding. Renders a 5 second clip in under 3 seconds.

The part nobody’s mentioning: H3 Max is a post-trained version of MiniMax H3. Chinese open weights model. fal grabbed the weights, retrained them on their own data, built their own inference stack around it, and came out with something better and faster than the original.

Try that with Veo. Try that with Sora. You can’t get inside a closed model, so you can’t do anything with it except pay for API calls.

And yet scroll LinkedIn and you’d think Claude and ChatGPT are the only two AI companies that exist.
They’re not, not even close. On OpenRouter, Chinese models went from a rounding error in late 2024 to the majority of top model traffic by early 2026. Qwen passed a billion downloads on Hugging Face this year. It’s already past 3 billion now. Four frontier grade open models shipped in a seven week window this spring: DeepSeek V4, MiniMax M3, Kimi K2.7-Code, GLM-5.2. That’s not a cadence, that’s a fire hose.

Why open weights keep winning:
You don’t need permission. fal didn’t ask MiniMax if they could build H3 Max. That’s the entire point of open weights, and it’s how a Chinese model ends up quietly powering a Western product.

Control beats cheap, eventually. A low price wins until it doesn’t. Open weights win because you host them yourself, fine-tune on your own data, and nobody deprecates the model out from under you at 2am. Companies that wouldn’t touch a Chinese API with a ten foot pole are running Chinese weights on their own servers right now. Once it’s a file on your disk, where it came from stops mattering as much as you’d think.

It doesn’t slow down. Blink and there’s a new frontier release out of some lab in Hangzhou or Beijing.

None of this means throw out Claude and GPT. A US government eval put DeepSeek V4 Pro about 8 months behind the leading American models. Claude’s still on top. But that same gap went from 20+ benchmark points down to under 3 in about two years, and China’s doing it on roughly a 23rd of what the US spent on private AI investment last year. Whether that line keeps bending the same way, nobody actually knows. But pretending it isn’t bending would be its own kind of dishonest.

We’re not sleeping on Chinese open-weight models. We’re just not hearing about them, because the people with the microphone are the ones selling closed ones.

H3 Max is the proof. A Chinese open-weights model, picked up by a Western company, tuned into the best video model on the market. Impossible to do if the weights are locked down.

That’s the story.

Try H3 Max here: https://lnkd.in/gGvp-A78

What's all the "buzz" about

Wed, 26 Aug 2026 12:00:00 GMT

As many of you know, I love playing around with new tech, so when I heard about Jack Dorsey's Buzz a few weeks ago, I was immediately intrigued.

If you haven't come across it yet, Buzz is essentially Block's take on Slack, but built from the ground up for agents. You can already talk to your agents inside Slack, but what makes Buzz different is that the agents can talk to each other... and any humans in the channel!

I've linked Hermes and Claude, my two most-used AI teammates, inside Buzz, and they're now running several projects together. What really caught me off guard is how they divide work autonomously and step in to help each other without being told to. They also automatically ask for my help when they need it or ask me to check on something they just did. Just like a real human teammate would do.

A couple of examples. Last weekend I asked Hermes (running on DeepSeek V4 Flash via Ollama Cloud) to handle a task and told it to loop in Claude if needed. It came back and said it didn't need the help, and just got it done. In another case, I dropped both Claude and Hermes into a channel and asked them to work on Tunepulse, one of my side projects. I watched them split the work, check in with each other, and deliver exactly what I asked for.

I've written before about agent orchestration being the real unlock here, and Buzz takes that a big step further. It genuinely feels like the missing piece, and the buzz around Buzz is well earned.

That said, it's still rough around the edges. The team at Block seems to be shipping fixes daily though, so it's definitely one to watch closely.

Check it out: https://buzz.xyz/

A note on AI and music.

Wed, 29 Jul 2026 12:00:00 GMT

The first email I read was from my label, asking why two of my tracks were flagged as AI on Traxsource.

Most of you know me from Adobe. Nineteen years, events team, keynote demos, generative AI activations. What shows up less on this feed is the other half of my life: I produce and DJ progressive and melodic house under the name Grish. Releases out, played out, the whole thing. It's not a side hustle I mention at parties, it's a second body of work.

I don't use AI to write music. No generated melodies, no generated arrangements, no generated drums, no generated vocals. I make these records on hardware, in my studio, slowly, and a human engineer mixes and masters them. That's the part I actually enjoy. Twisting knobs, dialing in a reverb until it sits right, building a synth preset from scratch, sitting with a vocalist and finding the take. Why would I automate the part I do this for?

I ran both tracks through SH Labs, one of the detection APIs Traxsource uses, and both came back as human made. The tag on my tracks names SoundPatrol as the source. So Traxsource uses two detectors and they don't agree with each other. One of them tagged my tracks, and now I'm the one assembling evidence to prove I made my own record.

That SoundPatrol name matters, because they are the ones gaining the most ground right now, with Universal and Sony both signed on.

To be clear about where I do use AI: artwork, promo images, photo cleanup, code. Given the day job, I'd be a hypocrite to pretend otherwise. But there's a difference between a tool that speeds up the work and a tool that does the work. In all of that, I'm still the one making the decisions.

Here's the part that concerns me beyond my own two releases. I've seen it suggested that highly repetitive arrangements have a higher chance of getting flagged. Repetition is not a symptom of machine generation in electronic music. It's the genre. A 32 bar loop that builds is the entire point.

To be fair, Traxsource did respond within 2 hours of me raising the issue and there is a real process, albeit a bit broken. I can't dispute this myself. They only accepts disputes from verified rights holders through their label system, so I now have to ask my label to ask their distributor to file on my behalf. I'm the only person in that chain who has the project files, and I'm the one person who can't submit them.

I welcome the transparency about AI in music. I want listeners to know what they're buying. But the methods used to flag it have to be rock solid, because a false positive costs an artist placement, income, and credibility they then have to win back.

Get the detection right, and I'm all for it.