__jf__

No, also because that would mean I have to read up on rope scaling or whatever to go beyond the default 256k context, which so far I’ve avoided. But you can allocate it as 220k context for 2 parallel Pi workers, that really speeds things up! Or as ~8x 50k context for dedicated parallel document processing.

Subtle quote from their website:

> Advancing the Frontier of Open Intelligence.

Too bad you probably need something like an 8x B300 node to run this. Black Friday is coming soon though, but I fear this open intelligence will be pacing anyone's liquidity frontier so I'll probably keep trailing the frontier for a while with nvidia/qwen3.8-27B-NVFP4 on dual RTX PRO 4000 in TP=2. I can almost fit 450k context in there.

Wow! TIL! I've been running a for loop around the two ordering variations to catch the winner of each turn and the difference is quite noticeable. In the options-after-body case in 47 of 100 attempts it classifies as phishing, whereas in the options-before-body case it classifies clearly as rickroll (94 out of 100 attempts)

Payroll sends you an email with a link to a Youtube video that plays a song.

Options after body:

    Average probabilities:
    Rickroll   0.5158 ( 51 wins)
    Phishing   0.4561 ( 47 wins)
    Spam       0.0281 (  2 wins)
    Joke       0.0000 (  0 wins)
    Legitimate 0.0000 (  0 wins)

Options before body:

    Average probabilities:
    Rickroll   0.9293 ( 94 wins)
    Joke       0.0549 (  5 wins)
    Phishing   0.0140 (  1 wins)
    Spam       0.0018 (  0 wins)
    Legitimate 0.0000 (  0 wins)

This was Gemma4-26B-A4B-NVFP4 by the way.

EDIT

Gemma4-12B-it-NVFP4 seems way less sensitive to option/body ordering:

Options after body:

    Average probabilities:
    Rickroll   0.9867 ( 99 wins)
    Phishing   0.0133 (  1 wins)
    Joke       0.0000 (  0 wins)
    Spam       0.0000 (  0 wins)
    Legitimate 0.0000 (  0 wins)

Options before body:

    Average probabilities:
    Rickroll   0.9401 ( 93 wins)
    Phishing   0.0336 (  3 wins)
    Spam       0.0250 (  4 wins)
    Joke       0.0010 (  0 wins)
    Legitimate 0.0002 (  0 wins)

Anyway, this for-looping stuff doing 100 calls to even a local VLLM API takes around 5 seconds in total, so this isn't anywhere close to sub-second Jev territory.

Gay Talese starts "The Kidnapping of Joe Bonanno" (Esquire, August 1971) with this:

Knowing that it is possible to see too much, most doormen in New York have developed an extraordinary sense of selective vision: they know what to see and what to ignore, when to be curious and when to be indolent—they are most often standing indoors, unaware, when there are accidents or arguments in front of their buildings, and they are usually in the street seeking taxicabs when burglars are escaping through the lobby. Although a doorman may disapprove of bribery and adultery, his back is invariably turned when the superintendent is handing money to the fire inspector or when a tenant whose wife is away escorts a young woman into the elevator—which is not to accuse the doorman of hypocrisy or cowardice but merely to suggest that his instinct for uninvolvement is very strong, and to speculate that doormen have perhaps learned through experience that nothing is to be gained by serving as a material witness to life’s unseemly sights or to the madness of the city. This being so, it was not surprising that on the night when the reputed Mafia chief, Joseph Bonanno, was grabbed by two gunmen in front of a luxury apartment house on Park Avenue near Thirty-sixth Street, shortly after midnight on a rainy Tuesday in October, the doorman was standing in the lobby talking to the elevator man and saw nothing.

Paul Strassmann wrote a book in 1990 called "Business Value of Computers" that showed that it matters where money on computers is spent. Only firms that spent it on their core business processes showed increased revenues whereas the ones that spent it on peripheral business processes didn't.

For vector generation I started using Meta-LLama-3-8B in april 2024 with Python and Transformers for each text chunk on an RTX-A6000. Wow that thing was fast but noisy and also burns 500W. So a year ago I switched to an M1 Ultra and only had to replace Transformers with Apple's MLX python library. Approximately the same speed but less heat and noise. The Llama model has 4k dimensions so at fp16 thats 8 kilobyte per chunk, which I store in a BLOB column in SQLite via numpy.save(). Between running on the RTX and M1 there is a very small difference in vector output but not enough for me to change retrieval results, regenerate the vectors or change to another LLM.

For retrieval I load all the vectors from the SQlite database into a numpy.array and hand it to FAISS. Faiss-gpu was impressively fast on the RTX6000 and faiss-cpu is slower on the M1 Ultra but still fast enough for my purposes (I'm firing a few queries per day, not per minute). For 5 million chunks memory usage is around 40 GB which both fit into the A6000 and easily fits into the 128GB of the M1 Ultra. It works, I'm happy.

From the pinned YT comment:

Artist Ostap Gordon just released a Russian yard in the winter in Unreal Engine 5.1 with high-quality assets created with Megascans and Blender, fully dynamic lighting with Lumen and Nanite for all objects. Check out this ultra realistic walk recorded with a RTX 4090 in a 4K monitor.

I ordered directly on Abbott's FreeStyle website.

I was primarily struck by the awesomeness of this whole autonomous glucose regulation thing. My last meal is usually around 18:30 in the evening and during the night glucose would fluctuate around 4.5 mmol/l between 4.0 and 5.0 in 1 hour periods, like a crappy PID controller that needs a firmware update. Other nights it would be flat instead of fluctuating, but unfortunately two weeks were too short for a controlled experiment, meal repeats and figuring out what caused the difference. Some nights it would show a couple of hypo's where glucose would drop to 3.5, quickly to be countered by an increase. I didn't notice a thing.

Additionally every morning before my alarm went off, I could see my glucose increasing, most likely preparing for wakeup, all by itself. Amazing!

It gave me a new-found respect for these otherwise invisible processes happening in this fleshy vessel on autopilot with closed cockpit doors. I only got to peek through a small window during 2 weeks.

Yep the FreeStyle Libre 2 is great. I bought one a few weeks ago to satisfy my curiosity. I'm a non-diabetic. It was around EUR 70 including shipping and lasted exactly 2 weeks. Sensor insertion was easy, quick and painless. 1-minute resolution with 8 hour memory between swipes. Curiosity satisfied!

Are there any perfectionists around here that have experience with Acceptance and Commitment Therapy (ACT)? It sounds like perfectionism-induced procrastination: it's better to give up than to play because playing may mean losing, and the chance of losing is unacceptable.

Near the end of “Understanding Michael Porter”, there is a Q&A with him that touches on this subject:

Q: “How do you do a five forces analysis if you’re an entrepreneur starting a new business in a completely new market space? Is strategy even relevant when there’s no existing industry or when conditions are still so fluid that there is no discernible industry structure and no direct competitors?”

Porter: “Strategy is relevant for any organization at any point in its trajectory. How to develop and sustain a competitive advantage is the core question that every organization has to answer if it’s to be successful and to prosper. In emerging industries there’s a lot of experimentation. What will the product ultimately look like? What will the distribution system look like? Will the product or service scope produce a stand-alone industry, or will this new idea become part of a larger or existing industry? There’s more uncertainty about the shape of things, but the five forces exercise is fundamentally the same with one big exception: instead of analyzing what already exists, you’re forecasting. And you probably know quite a lot about all of the five forces but one. You know the customers you’re targeting. Are they likely to be price sensitive? You know who your suppliers are or who they are likely to be. How powerful will they become? You know the substitutes and can identify the likely entry barriers. What you don’t have yet are actual rivals. That’s where you need to think through who those might be. Will the rivals most likely come from adjacent industries? Or from companies that already exist in other countries? Or will the likely rivals be new start-ups? How would each of these rivals be likely to compete? So even when you’re inventing new market market space, you probably already know more about the five forces than you realize. Doing such analysis is important because if you’re creating something that’s truly valuable, don’t kid yourself that no one will follow you. There is no such thing as a market where competition is irrelevant, as nice as that might sound. The idea that innovation allows you to ignore competition is a fairy tale. So you have to have a hypothesis for how the industry might take shape once there is an industry. Early on, there are many paths the evolution can take, many choices you can make that will have an important impact on how attractive the industry will become. Decisions you and others make over time will begin to lock in the basic economics, making industry structure less fluid. So it’s crucial to see different paths for how the industry might evolve, and to ask the basic questions about the five forces, so that you can make choices that will put the industry on the best possible path.“