Still Doing That Thing…

I’ve been ranting a bit here over the last few months, talking about my stance(s) on AI in all its myriad forms. For a look back here’s Part 1 and Part 2.

The things I keep running into in my day job, weekend writing gig, and nighttime doomscroll are about as consistent as a metronome. The authors in trad publishing are dead against all forms of LLM use. The publishers in the same realm are for it exactly as far as they can push their author cash cows. Meanwhile, the indie world and my local author scene are desperate and willing to try anything to get a few more ebook sales. Fledgeling writers getting their start on their first novel or memoir aren’t extremely online enough to even know how they’re supposed to feel about LLMs. Though, the kids in school are hating it.

All of the above are using LLMs whether they admit it or not, even to themselves. Meanwhile, the corporate folks are starting to wake up to the fact that no, they can’t lay everyone off and do it with ChatGPT. Tokens cost just as much or more money than people (especially when you don’t know what you’re doing with them), and BIG SURPRISE, the people who know what they’re doing, who know how (to write, read code, do analytics, talk to people over the phone, etc.) not only they are worth the money, they can get even more done if you give them tools to do their jobs more efficiently and focus on what they do the best. (But pssst… Those tools include LLMs too)

As I type this, the first big popping sounds are coming from Wall St. Anthropic’s threatened IPO is rumored to be rated as Junk before it launches. Microsoft offered buyouts to 1/3rd of their worldwide workforce to offset their debt. Intel’s taking a beating, since no one can afford PCs. Google was half-honest in their quarterly earnings and said it’s going to be half a decade before they can start making money again. Oh and the world’s biggest bank heist in history just went down as a footnote in the day’s news cycle as they think N Korea just made off with $3-6 Billion in crypto.

It’s wild, and like a metronome, it swings from one end to the other and back again but all sounds the same.

So back to my thing… There’s one reason I haven’t talked more about it. Well, one reason in addition to all the chaos noted above. That’s a lack of models that check all the boxes. I can be sure everything in my editor system work well, as designed and built. I’ve gone over the whole stack by hand. It’s secure, scalable, and sound. I can assure most of my checklist.

• Author Privacy – Check
• Vetted Models – Check
• Editorial Quality – Check
• Ethical Data Sourcing… ummm

My early tests were all done with Claude, and Opus still holds the quality benchmark record, but ethical? Yeah, no.

On the local Open Source model front, the best ones are coming out of the Chinese labs. Qwen, Kimi, Deepseek… All ethical trainwrecks.

Some recent happenings on the EU side of the pond are a glimmer of hope, however.

I should say a very slight glimmer. Just because a company is able to check the boxes on EU regulations, doesn’t mean they’re ethical or has anyone or anything other than their bottom line in mind.

To name a few, Swiss AI’s Apertus models are doing good work in keeping their datasets responsibly sourced and trained. The German group Soofi is building everything on the up and up, even documenting the power and water sources of their training datacenters. These are also filling a gap of multilingual models needed to serve communities across the EU, so props to all the folks working on them.

While these models are theoretically checking the boxes on the ethics front, they are very academic. Soofi in particular is holding their models back from wide release while they test with trusted groups. Apertus is available now and on my list to test as a local model, but if the score cards are an indicator, I know already that I’m going to be disappointed in its performance.

That leaves Mistral. These folks are a perennial runner-up when it comes to frontier model building, but they’re the only one up there checking the EU boxes, so I have to give them lots of credit. You can read between the lines of their filings though. They acknowledge a lot of their baseline training was done before the 2025 EU rules. They can’t guarantee the IP on those early datasets, and they’re not likely to start from scratch. They’re open and transparent about everything since, however. They also have to abide by GDPR, and their partner datacenters are also in the EU, under their environmental regulation.

I’m leaning a lot on the EU here, but someone has to do the work to keep these folks in line. I don’t see the US or China even trying, and everyone else in the world is just itching to grab one AI headline. (I’m looking at you, India)

SO, that leaves me in a bit of a limbo. I could release my editor today with a Claude back-end. I could give contract-tight privacy guarantees. I know the output would wow people, but I’d feel shitty about it, given the runaround Anthropic has handed the author community with their settlement fiasco. (At least they’re voluntarily paying though. That’s worth something, right?) I’d also be contributing to the rat race datacenter build out and all the fallout that’s gonna cause in the coming months.

I could release with a local machine model, keep the wattage low, and give total data assurance that author work never leaves my machines, save for some encrypted cloud storage. But it’d run a Qwen model from Alibaba, who doesn’t even bother saying they tried to source the data ethically. They just say, “Take it up with the Chairman.” Copyright? We don’t need no stinking copyright. We’ll even train our model by subscribing to the US models and reinforcement learn through straight up copypasta.

That brings me back to Mistral. I can run their smaller models. The quality is meh, but not much worse than Qwen. The ethics and environmental impact are as good as the EU can enforce right now. I can give data assurance locally or in their cloud. Mistral Large 4 might even meet the quality bar.

So am I talking myself into releasing this thing (if Mistral Large 4 is any good, which has yet to be seen)? This will be the first potential batch of tests that have a chance.

Wish me luck? I guess?