Remco

Turning AI agents into tools that actually build and ship.
Building:
LuminaReviewer
Lunchlink
Starryframe

Data scientist · NL.

[Remco @RemcoAI]

I spend most of my time building with AI and figuring out what is actually useful.

Not just asking models random things, but turning them into tools and workflows that help me build real products and ship faster.

Right now I’m building products, testing agents, comparing models, and sharing what actually works and what doesn’t.

If you like following someone who is actually trying to make AI useful in the real world instead of just talking about it, follow along.

[Remco @RemcoAI]

@themishra4402 Generating images uses a lot of compute and will burn down the servers

[Remco @RemcoAI]

@trikcode You just scroll trough you X feed

[Remco @RemcoAI]

@MeloxJourney You got 3 active subscriptions 🎉

[Remco @RemcoAI]

@T_Zahil Add a message gateway like telegram, discord or whatsapp and just talk with it. It can do action at a higher autonomy

[Remco @RemcoAI]

@EU_Commission Are you kidding?? You cant be serious about this. There are not many tech leading companies left in the EU and those who are in the EU consider leaving...

[Remco @RemcoAI]

@Akasheth_ Its blazing fast and performs quite well. I am currently building a web application with it and havent had any problems yet

[Remco @RemcoAI]

@Google @googlegemma Start intoducing Gemini Pro 3.5

[Remco @RemcoAI]

“Mythos find all ways to increase tax and regulations and make Europe green again without starting a revolution. Dont worry about multinationals leaving the continent.
Make no mistakes”

[Polymarket @Polymarket]

JUST IN: Anthropic offers EU access to Claude Mythos.

[Remco @RemcoAI]

@shub0414 Based on your question this token had the highest predictability

[Remco @RemcoAI]

Marvell closing +32.5% at $291 puts it around 100x trailing EPS and 48x forward EPS.

That is rare situation for a scaled semiconductor company. Nvidia is around 18x forward earnings, Broadcom 26x, and AMD/Arista roughly 40x. So the market is no longer valuing MRVL like a normal networking chip supplier.

The bet is that AI clusters need far more networking, optical interconnect, custom silicon and data movement and that Marvell becomes one of the scarce picks-and-shovels winners.

But the valuation also means the bar is very high. At today’s price, MRVL would need roughly $14.5 EPS just to look like a 20x earnings stock, versus $2.91 trailing EPS and $6.11 forward.

That does not mean it has to crash. Nvidia also looked expensive before earnings exploded.

But this situation is different: Nvidia’s multiple compressed as profits scaled. Marvell now has to prove that AI networking and custom silicon can create the same kind of earnings catch-up. They will need to have a similar monopoly position as Nvidia has.

[Remco @RemcoAI]

AMD’s 192GB unified-memory Ryzen AI Max is interesting because it changes what “local AI” can mean on a normal desktop.

The practical translation: 70B-class models like Llama 3.1 70B, Qwen 72B or DeepSeek-R1 Distill 70B become realistic local workloads at 4/8-bit quantization, with room left for context, tools and agent loops.

It probably won’t run a full GLM-5.1 model in any practical way 750B total parameters is still far too large once you include quantization overhead and KV cache.

But that’s not the point. The point is that local AI moves from “small assistant demo” to “serious private model running on your desktop”. A world where intelligence is in your desktop at home is becoming more realistic every month.

[Remco @RemcoAI]

Today I used Hermes with Grok Build to run a full outbound workflow for LuminaReviewer: collect and verify 200 research/evidence-synthesis contacts, dedupe against previous sends, remove an excluded segment, send a test email, then send the email campaign in 8 batches with logging.

The interesting part is not the email sending. It’s that agents become useful in reliably handling boring stuff for us.

Maybe it should make a post for me on producthunt too

[Remco @RemcoAI]

Ok the limits were clearly short term. I have been heavily using Grok Build today and am at 5% montly usage. At this tempo i wont make it untill end of the month, but at this point is seems to be worth the 35€.

[Remco @RemcoAI]

Grok Build hit rate limits really really fast in Hermes agent. Its now only returing the following message: 'The model provider is rate-limiting requests. Please wait a moment and try again.'

After waiting for 30 minutes still the same message appears. I also cant check the usage in my XAI console. Only usage with the API key is visible in there.

@xai At least fix that we can see usage and limits in the console so we know how long the rate-limits last.

[Remco @RemcoAI]

@neerajjj6785 I just did. I got rate limited in less than an hour. So people are using it but only for short periods of time

[Remco @RemcoAI]

Grok Build hit rate limits really really fast in Hermes agent. Its now only returing the following message: 'The model provider is rate-limiting requests. Please wait a moment and try again.'

After waiting for 30 minutes still the same message appears. I also cant check the usage in my XAI console. Only usage with the API key is visible in there.

@xai At least fix that we can see usage and limits in the console so we know how long the rate-limits last.

[Remco @RemcoAI]

@chribjel The qouta is too limited to be usefull.

[Remco @RemcoAI]

@DeepakNesss Simply because it runs in the Colossus

[Remco @RemcoAI]

Composer 2.5 in Grok Build is super fast and works fine for straight forward tasks. I'll test it out later with more complex questions. I really really like the speed.

Great work from XAI and Cursor

[Remco @RemcoAI]

Composer 2.5 is now available in Grok CLI. Lets check out how it performs!

[Remco @RemcoAI]

First impression of Grok Build 0.1: already a clear step up from Grok 4.3 for agentic coding workflows.

Grok 4.3 was surprisingly weak at tool calling in my setup. Build 0.1 seems much more reliable so far, which matters more than raw benchmark scores once you put a model at work in Hermes where toolcalling matters.

Will stress-test it on more complex tasks next. Ill keep ya updated