Patricio Lobos

openbi.no now is on average 4x to 9x faster than any other popular BI tool, completly destroying PowerBI , specially on Mac Silicon were pulls ahead 9x. It is -in simpliest terms- the fastest in-memory analytics database, than anybody can download right now.

Had a meeting with @InnovasjonNorge, a state owned "start-up incubator", their pitch was literally this: "Give me USD 400 a month, you can use our offices a few days a week, we help you with a deck of slides" I have use already from my own pocket USD 50.000, I have run million dollars projects in banking and finances, and I have 2 jobs.... like these ppl are completely fucking clueless. @Norge and the EU and going straight to medieval times, @1x_tech had to GTFO here.. there is no future , only refugees and taxes....

I went to talk to one of the largest supermarket chains the other day, they where using F128, F64 and F16 capacities in Fabric, I offer them to save USD 10.000 per month by checking 1000s of dataset, lineages, etc. Jon, the leader of reporting that describes himself in linkedin with the headline "I want to help organizations replace manual and cumbersome processes with solutions that simplify everyday work." Refusde the offer with the following line "This will be boring for you". It remind me the head of a financial institution that told me "But with what you show me I will be the VP of 5 ppl not 100". @chamath says that AI is not producing outcomes. It is because the C-management need to get rid of the meat proxy, but these employees are not for the company or customer's benefit, but their own, their networks of favors, etc. Most leaders are simply clueless of what they actually do, an with AI they are even more clueless about how to apply it to...

using PRMG for our website querex.no Procedural real-time motion graphics, if you don't know what is it take a look (2Dcanvas and JS) I believe the internet is going to be on fire with this, even @3blue1brown will use it....

"LLMs have left the chat" The last two years of “agents” still ran on a 2018 interface: think in vectors, speak in paragraphs. One model wrote a briefing. Another read it. Code scraped the briefing for a JSON blob. Every hop paid the same tax: decode tokens, lose nuance, wait. C2C and Jev are the first widely noticed attempts to walk away from that tax in opposite directions. Together they say the same thing: language was a compatibility layer, not the native protocol. Key insight : We need to get rid of the meat proxy C2C, from Tsinghua and collaborators (ICLR 2026), cuts the conversation between models. Two LLMs prefill the same context, then a trained fuser projects the Sharer’s KV-cache into the Receiver’s space, gates it layer by layer, and residual-adds it. The Receiver decodes from the fused cache. No intermediate sentence is ever written for the hand-off. That is why the paper can claim roughly 2–2.5× lower latency and a few points of accuracy over text-to-text agent loops: the features that would have died in a paragraph survive as tensors. The cost is pair-specific adapters, a same-context assumption later work already pokes at, and a channel humans cannot read. Key insight "Telephaty between models" Jev, TypeSafe’s System One model (Sept 2026), cuts the conversation between model and software. You send state plus a closed question set : pick one option, score a rubric, answer true/false and get calibrated probabilities in one parallel pass. There is no prose, so there is no malformed JSON and no invented category. RLCD is supposed to make “0.8” mean something a if statement can trust. Vendor numbers put latency in the 70–500 ms band and price at $0.042 / M input tokens with free output; the 200× / 400× headlines are company highs, not a settled benchmark. Architecture is still closed. The product is not a chatbot. It is a reflex. The synthesis of both is a stack, not a LLM replacement. Generative models still write plans, code, and explanations. C2C is how those models share internal state when they sit on the same box and text would be both slow and lossy. Jev is how the loop decides whether to call them: route, gate, score, escalate. One removes speech between minds. The other removes speech between mind and machine. That is what “left the chat” actually means. The chat window is still there for us. Inside the system, the interesting traffic is starting to look like caches, gates, and typed distributions channels that never needed a sentence in the first place.

I'm using @deepseek_ai V4.1 Flash for all analytics now, finally a model that does what is told to do, i.e. is not waste of time like Opus 5 @AnthropicAI b.s. bloatware. If you need your CEO, L1, L2, L3 to have "actual AI" agents runnings endlesly, full Quant, Machine learning within seconds i a lively QA over @PowerBI, @Qlikview, @salesforce or any other source, in your IPAD, iphone or PC, there is nothing that comes even close on price, intelligence and speed.

When are we going to have a @X for the @AppleTV @elonmusk @grok ??

Running @Alibaba_Qwen 3.8 Flash (262K context) in a Windows Server 2025 from 2014!!! With 256GB Ram and a RTX blackwell Pro 4000. 1.57 tokens per second, @lmstudio 100GB RAM usage…. Amazing!!

"It becomes self-aware at 2:14am Eastern Time, August 29th." do we get a reset @thsottiaux to celebrate Skynet?

if you think you can run a chinese model against @PowerBI you can at least be sure to look away from @deepseek_ai it has zero understanding or notion of what DAX is, also behaves poorly with MCP. @AnthropicAI takes a few seconds to get a DAX ready, @openai many a couple of seconds more. Why this matters? 40 millions users and all top 500 Forbes companies use PowerBI, so in the context of "buying your own machine / sovereing AI for banking, finance, etc" forget about it!