Daily AI Fresh News · October 31 Evening Edition: DeepSeek's Open-Source Double Whammy

Lots of new stuff in AI today, and the evening wave is pretty lively.

Today we're covering a few things: first, five hands-on projects for an AI portfolio; second, making videos with Claude Code; third, a head-to-head test of eight video models; fourth, budget-friendly fine-tuning for students; fifth, building an AI SaaS with Next.js.

And then two big ones: DeepSeek^Ad's open-source model is back, and the mobile version of DeepSeek Harness.

Yeah, I've been keeping an eye on those two as well.

So let's start with the news roundup and go through them one by one.

Just saw a video on YouTube about how to build an AI engineering portfolio that can actually land you a job.

It gives five free projects: a systematically evaluated RAG, a local small-model benchmark comparison, an MCP server, a serious fine-tuning attempt, and a cost-latency dashboard.

I think that approach is right. The point isn't how many projects you have—it's showing how you make decisions.

Two or three done deep beats five done shallow, way more. Recruiters want to see how you handle constraints and how you shore up testing.

Right, and the video also says to honestly write down your process, hardware and software versions, and even your shortcomings.

Second item, also YouTube: someone used just one voiceover track and had Claude Code write an entire SaaS promo video as code.

I saw that one. The key is that animation events are bound to each word, not estimated by frame count.

So when the voiceover changes, you don't have to realign keyframes. That's the real difference from manual editing.

He said delivery time went from weeks down to a day, and that the approach can be reused for any product.

But the video didn't disclose the model version, prompts, or failure rate, so don't buy it all just yet.

Storyboarding, visual specs, brand consistency—humans still have to decide those. It saves execution, not judgment.

Fair point. Third item, again YouTube: someone took the same complex continuous-shot prompt and blind-tested eight AI video models.

Kling 4.0 took first place alone with 35 points, then three models tied at 33, and a bunch in the 29 to 32 range.

Three takeaways worth noting. Complex prompts are a common weak spot for everyone, and image quality isn't the deciding factor.

Also, a version bump doesn't mean across-the-board improvement. Seedance 2.5 looks better but actually missed some instruction-following.

That conclusion sounds a bit counterintuitive—a higher version number could actually be a step back?

Totally normal. Model optimization usually involves trade-offs. Tune one dimension and another might loosen up.

OK, fourth item: a Reddit post about doing model training and fine-tuning on a tight budget.

Three core ideas: freeze the base and only train PEFT like LoRA, use lightweight backbones for CV, and add memory-saving tricks like mixed precision and gradient accumulation.

Right, you can get started on a single consumer GPU. But the original post didn't give hardware or cost data, so don't treat those numbers as conclusions.

Last item: a full-stack walkthrough on YouTube building an AI SaaS with Next.js 16, OpenAI, and Drizzle.

It sorts out auth, the database, free quotas, and a demo mode that auto-degrades when there's no API key.

The value of that project really isn't the amount of code—it's that it cleanly handles the trickiest spots like session security and cost control.

Alright, that's the news for now. Next, let's dig into tonight's hottest topic: DeepSeek.

There's a video on Bilibili with a pretty bold title: DeepSeek crushes the field, the king of open-source models returns.

Your feeling is right—the buzz this time is really high. I saw comments saying the open-source camp is flexing again.

So what exactly makes it strong? High benchmark scores, or just smooth to use in practice?

That's not the point. The point is the act of open-sourcing itself. No matter how strong a closed model is, you can't tune it or change it.

Once an open model catches up to the top tier, enterprises get a fully controllable path. That psychological impact is bigger than the scores.

But for open models to be deployed, who eats the hardware and tuning costs? Aren't we back to square one?

That's why you look at them together. Even students can fine-tune on a single card now—the barrier is coming down.

Open source plus low-cost fine-tuning—those two trends stacking up is what's really exciting this time.

Got it. The second one is even bolder: DSH Go, the mobile version of DeepSeek Harness—AI takes over your phone.

Several reviews on Bilibili are getting a lot of traction, saying it can turn your phone into something like the Doubao phone experience.

I think this direction has more imagination than model scores. The phone is the closest layer between ordinary people and AI.

Note the word Harness—it's not just a chat box, it's more like a shell that can operate the phone.

Wait, you said taking over the phone—what about permissions and security? I wouldn't just hand that over.

That's the key tension. The stronger the capability, the more you need boundaries. Reviews are hot now, but real adoption depends on how the permission model is designed.

So you see, on one side open models are pushing costs down, and on the other AI is being stuffed into the phone as an entry point.

Right, those two are really two sides of the same story. The bottom layer is getting more open, and the top layer is getting more personal.

To sum up: DeepSeek's open-source return plus the mobile Harness shows the competition has shifted from who scores higher to who's more usable and controllable.

Exactly. Whoever nails both capability and boundaries is the one who laughs last.

Alright, that's it for today. That was a lot to take in—take your time digesting it. See you tomorrow.