Patricio Lobos

“I cannot fiddle, but I can make a BI system into an insight factory.”

[Patricio Lobos @strandedinoslo]

@MacRumors 1TB or nothing, built from the ground up to be stacked in 2,4,8 units, else is just useless if you want to run Chinese frontier models

[Patricio Lobos @strandedinoslo]

@SaraDiscovers Statisticians / Data Scientists doing that 11 years ago

[Patricio Lobos @strandedinoslo]

My clients range from CEOs to technical teams. With Querex MCP for Power BI, they can query anything extremely fast in an atomic, non-destructive way, whether it is DAX, the underlying SQL, M code, or, when there is doubt, the exact value in the core table in Fabric, on-prem SQL, or SQL Server Analysis Services.

This is why MCP is so powerful. We don't expect a CEO to run CLI commands or Python scripts. They simply want Grok, Codex, Claude, or Microsoft 365 Copilot to give them the answer within seconds.

In our experience, Claude is currently the fastest, usually responding in under three seconds on average.

[Patricio Lobos @strandedinoslo]

@bcherny here you can see both @OpenAI codex and @github copilot rendering images... so it is only a problem with both @claudeai and @grok

[Patricio Lobos @strandedinoslo]

@bcherny found a small but real MCP gap in the @claudeai desktop/Max app.

My MCP server returns one tool result containing: text + an inline image/png block annotated { "audience": ["user"], "priority": 1.0 } + a download URL. Fully spec-correct.

What happens in Claude:
• the runtime decodes the image into the model's context, the model sees it.
• but the app never surfaces that tool_result image into the human transcript
• and it doesn't expose the base64 back to the model, so the model can't re-route it via present_files either.
So a tool_result image explicitly tagged audience:["user"] reaches the model and dead-ends. The human only ever gets text + a link.
Per the MCP spec, audience:["user"] is a routing instruction to the client to show that content to the user.

Honoring it (promoting tool_result images into the transcript) closes the loop. (Diagnosed by Claude itself mid-session.)

[Patricio Lobos @strandedinoslo]

@UnitreeRobotics dancing again!!! 🤮, please show us a robot that actually do something useful, like support MCP, being able to be recognized as a mouse and keyboard, dexterous hands, etc....

[Patricio Lobos @strandedinoslo]

@nikitabier two things are killing Grok for real MCP work:

It makes me re-enter MCP auth every single day. No persistent session, no token refresh — full re-OAuth daily like it's 2010. (Mobile app do not support MCP)

It can't render images returned by MCP tools. My server hands Grok a valid base64 image/png content block (straight from the MCP spec) — Grok drops it and shows a download link instead. Claude Desktop, VS Code, and ChatGPT all render it inline. Grok just ignores it.

The image data is right there in the tool response. Fix image content blocks + persist the auth and Grok actually becomes usable for tooling.

[Patricio Lobos @strandedinoslo]

It will burst only if @jietang releases a FABLE-level model in the next 70 days. Then all closed American companies will have no moat at all. Still, hardware is king: a Supermicro with 8x GPU AMD 350 is 400K at least.

So once the closed labs go down, the next bubble will be @Huawei running a Chinese FABLE model in a rack that you can order on Alibaba for 200K.

[Patricio Lobos @strandedinoslo]

@barrowjoseph @modal @charles_irl @MistralAI you need to make your prices cheaper. You are currently chargind 2 x 1000, can you keep the old USD 1 per 1000 pages?

[Patricio Lobos @strandedinoslo]

78 days between version 5.1 and 5.2 release, if RSI (Recursive selv-improvement) is true now, i guess next version in 60 days and we will have Fable level capabilities and the biggest bubble burst in the history on the American stock market.

I hope they release it with a "guide" to Chinese hardware to run, if they are competitive to Supermicro 8 AMD MI355 or MI400, then we will order such machines...

[Patricio Lobos @strandedinoslo]

Windows Fax and Scan was remove from the latest update, so now goes to optional features, the whole windows experience now is like walking in an indian street, being harrased to use copilot, onedrive, screaming banners, pop-ups etc and I'm using Windows Pro 11.... @Windows is literally forcing users to move to Linux like @LinusTech said https://t.co/NQycGjqtiZ

[Patricio Lobos @strandedinoslo]

@omarsar0 In some areas we still are very far from that a single synthesis in FPGA can take easy 3 or 4 hours, then test and back, current models (exception Fabel class) are nowhere near to code on FPGA and Fabel was blocking dev on FPGAs anyway.

[Patricio Lobos @strandedinoslo]

@__tinygrad__ We might buy one Supermicro AS -4126GS-NMR-LCC | 4U GPU Server (DP AMD 4U Liquid-Cooled System with AMD Instinct™ MI355X 8-GPU)

According to specs it runs at 45 - 50 db... if you can share how you doing with KIMI code 2.7 👌👌👀👀

[Patricio Lobos @strandedinoslo]

@skirano They used @elder_plinius jailbreak… 🫡

[Patricio Lobos @strandedinoslo]

@__tinygrad__ To run Kimi 2.7 you need AMD MI355 x 4 at least

[Patricio Lobos @strandedinoslo]

@klingefjord Calling Iceland for that one 😃

[Patricio Lobos @strandedinoslo]

@NandoDF It is not hidding in Norway, I can tell you that!

[Patricio Lobos @strandedinoslo]

@ClementDelangue FPGA dev (Vivado CLI) is banned

[Patricio Lobos @strandedinoslo]

@simonw @edzitron for most queries yes, sadly for biology, FPGA domain, ML research there is no "fable" and if @sama models can actually reply those queries, then no war price is needed.