clarissa lexber

Ex-Big Tech | Now exploring AI full-time
Models, Agents, AI tools & open-source projects
Sharing what I actually test, use, and find worth talking about

[clarissa lexber @ClarissaLexber]

@AnaMiles245714 hahaha😂

[clarissa lexber @ClarissaLexber]

Spent quite a while digging around, only to realize DeepSeek once again didn’t make any announcement. DeepSeek-V4-Flash-Vision-Exp just suddenly appeared on Hugging Face Another very DeepSeek-style release.

Looking at tests from overseas developers, given the model’s performance, it doesn’t seem particularly good at handling dense documents, small-font tables, or fine-grained charts. But it should be fine for reading normal screenshots and getting a general sense of a design.

Then again, it isn’t really meant for general-purpose visual understanding. It’s mainly there to give Agents an extra source of visual input — almost like DeepSeek is catching up on some unfinished homework.

Most people probably won’t have the hardware to deploy it locally, so API access will be the only realistic option.

At current pricing, V4-Flash-Vision-Exp is currently the most expensive of the three low-cost multimodal models from major Chinese providers, and it’s also the only one that isn’t natively multimodal…

[clarissa lexber @ClarissaLexber]

@FokLouis144 yes

[clarissa lexber @ClarissaLexber]

@hutun1314 Feels like DeepSeek always likes to make its moves quietly, but the practical potential of this update still looks pretty promising.

[clarissa lexber @ClarissaLexber]

@yjwds4567 Yeah, that’s very on-brand for DeepSeek. No hype, no countdown — the model just shows up and people start testing it.

[clarissa lexber @ClarissaLexber]

This comparison showed me the clear advantage of long reasoning in body proportions and line work,
while also making me anxious about the nearly 20-minute wait.

SVG generation remains a hardcore task where quality and efficiency are still hard to have both.

[Harshith @HarshithLucky3]

GPT Astra max vs Gemini 3.8 Flash high

my BMW M4 CS side view SVG test is still not solved

~19 min vs ~42 sec

Astra result is the best so far

but look at the speed difference with Gemini 3.8 Flash

[clarissa lexber @ClarissaLexber]

@JinjingLiang @orca_build Your Macs didn’t burn tokens, they burned calories.
AWS for pain, Hetzner for sanity.
Or just summon more laptops like Pokémon.

[clarissa lexber @ClarissaLexber]

I’d probably cancel Plus for a month and see what happens.

Not because the product is bad, but because the pricing/limit logic feels too unpredictable right now. AI tools are moving fast, and locking yourself into one ecosystem only makes sense when the experience feels stable.

Right now, it doesn’t.

[clarissa lexber @ClarissaLexber]

Codex has been frontend/backend separated for quite a while now, so in theory building a GUI for it shouldn’t be that hard.

But after looking around, I still couldn’t find a Codex web GUI I was actually happy with, so I figured: why not just let Codex see if it can build one itself?

One problem with some of the existing GUIs is that if they don’t keep up with Codex updates, things like the model selector on the web side can get out of sync with Codex itself, and then the whole thing just gets stuck or becomes unusable.

And the more features a GUI tries to pack in, the more complicated the code gets — which usually means more bugs too.

I wasn’t very happy with the first version Codex made, so later I asked it whether there was some ready-made web design we could just reuse. That way we wouldn’t have to waste time messing with GUI details — just wire everything up and call it a day.

Then it discovered that the latest version of Codex has some weird --remote feature…

And it basically just connected the TUI directly into the webpage for me.

I mean, it’s technically not wrong.

It’s 100% feature-complete and perfectly aligned with Codex CLI.

Because… it literally is Codex CLI.

But dude…

The whole point of a web interface is cross-device access.

How exactly am I supposed to press up, down, left, and right on my phone?

[clarissa lexber @ClarissaLexber]

@A_taozhu This is exactly what spare machines are born for, ah.

No fear, no guilt, just pure chaos engineering.

[clarissa lexber @ClarissaLexber]

Using AI to deploy AI, then using AI to write AI inference software.

Since this is just a Strix Halo test machine sitting around unused at the company, I had absolutely zero psychological burden.

Straight into visudo for passwordless sudo, then --dangerously-skip-permissions.

Now it’s officially AI all the way down.

That said, there’s still a funny contradiction when it comes to running local AI on Strix Halo.

If you’re not very familiar with AI / Agent tools, getting ROCm working can be painful. Without an Agent helping you, every ROCm version, model, and quantization format can turn into another round of hunting for the right compiler and toolchain.

And if you’re already skilled enough to use AI to wrestle ROCm into submission, chances are you probably don’t care that much about the limited compute Strix Halo offers in the first place.

So maybe the perfect Strix Halo user is someone like me:

using it mostly as a toy, where the real fun is not the inference performance itself, but the process of making ROCm behave.

[clarissa lexber @ClarissaLexber]

@wangfenganc Exactly 😂 Getting ROCm to behave is basically the benchmark at this point.

[clarissa lexber @ClarissaLexber]

@CryptoGouba Exactly 😂 At some point, fighting ROCm with AI becomes the actual hobby.

[clarissa lexber @ClarissaLexber]

This feels like a capacity routing issue rather than a true “reset” problem. The quota may have refreshed, but the specific model pool you selected can still be temporarily saturated.

I’ve seen this happen when demand spikes after resets. Switching models works, but it’s frustrating when you’re in the middle of a flow.

Would be great if Codex showed clearer status: quota limit vs model capacity vs temporary outage. Those are very different problems.

[clarissa lexber @ClarissaLexber]

@Ananth7e Classic “20x” until you read the tiny math, ha.

Feels less like pricing and more like a trap.

[clarissa lexber @ClarissaLexber]

@FlowOpsDaily Yeah, that’s exactly what I was going for. Let the demo speak for itself without adding too much explanation.

[clarissa lexber @ClarissaLexber]

https://t.co/fylj01I73i

[clarissa lexber @ClarissaLexber]

@kingzw888 Yeah, especially with better optimization around Apple Silicon. Local models are becoming much more practical now.

[clarissa lexber @ClarissaLexber]

@Vesper2536 Yeah, that’s one of the best parts. The hardware requirements are pretty clear, and even the entry-level Macs can already run some useful local models.

[clarissa lexber @ClarissaLexber]

@gabrielchua This is exactly the missing piece ah.
Tiny ideas can finally ship without turning into a whole infrastructure project.

[clarissa lexber @ClarissaLexber]

@proven018 I’ve been looking into WorkBuddy too lately!