RandyRanderson

I learnt the facts of life ... from watching The Facts of Life

I think fb was one of the pioneers here. You would set the privacy settings and then they’d redesign the settings and coincidentally they would be just a little different and all be maximum share by default.

One of the reasons I stopped using it.

My claim is that this is due to alignment

Expressing charts in XML, JSON, etc is a terrible idea. There are several existing examples of this.

Ppl who think this is a good idea should be compelled to maintain some M4 or other configuration "files". These configuration files ALWAYS devolve into a full language if they are actually used. So now you have the language you wrote your app in and then you have another, shittier language for your configs ... that you have to maintain.

Please stop doing this.

What should we do? In this case just express the chart in a natural language. Where there are uncertainties, the LLM should ask questions to clarify and record the results.

It's import to solve the meta-issues here:

1 you didn't google current events before asking the first Q. Try that before asking someone else to.

2 mid-tier LLMs are capable of reasoning through your second Q. Go ahead try that next time.

3 you haven't accepted that 2 is true and the ramifications thereof.

In Canada this is a huge scam. The government advertizes that it's funding incubators. Great, right?

The money doesn't go to the start ups - it all goes to large tech companies like IBM, etc, because, obviously, IBM knows about innovation.

The cover is that the government doesn't know tech so it will give money to trusted partners and they will choose who to give the money to because they've been doing such a great job innovating in Canada. Surprise: they gave the money to themselves!

You might have wondered why all these incubators in the crypto era were desperate to get you to go to their office. You might have also wondered: what fool is paying for this nice office in downtown Toronto where the prices are crazy-high? The taxpayer.

All of that money was completely wasted and worse, little of it went to actual start ups.

Randy's Axiom 173: VERY widely used techs are some of the worst, ill-conceived techs ever.

However, no one will admit this because:

. many people don't know them (the techs) well enough because they're so abstruse. Some people would argue this a necessary feature.

. many people only know them and so can't compare them to something else

. people that do understand them typically have financial or ego reasons to look the other way

. people that do understand them AND are rich and happy are afraid of all the other ppl in this list.

Generalization: any widely use tech must be extremely complex or extremely simple (eg TCP).

Canada here: it can absolutely, definitely get worse.

Why is it surprising that, at some point, more information will lead to worse performance?

It seems obvious. Moreover, in a simple model, it seems like whatever tokens you do add have to have MORE information than the average in the existing window.

In a non-trivial model (and this is the model I would choose), since you are adding them to the end, they likely have to have MUCH more information.

Proof as always is an exercise to the reader.

Fable is 2x latest Opus:

  ┌─────────────────┬──────────────┬───────────────┬────────────────────┬──────────────────────┐
  
  │ Model           │ Input ($/MTok)│ Output ($/MTok)│ Batch Input (−50%) │ Batch Output (−50%)│
  
  ├─────────────────┼──────────────┼───────────────┼────────────────────┼──────────────────────┤
  
  │ Haiku 4.5       │    $1.00     │     $5.00     │       $0.50        │        $2.50         │
  
  │ Sonnet 4.6      │    $3.00     │    $15.00     │       $1.50        │        $7.50         │
  
  │ Opus 4.7        │    $5.00     │    $25.00     │       $2.50        │       $12.50         │
  
  │ Opus 4.8        │    $5.00     │    $25.00     │       $2.50        │       $12.50         │
  
  │ Fable 5         │   $10.00     │    $50.00     │       $5.00        │       $25.00         │
  
  └─────────────────┴──────────────┴───────────────┴────────────────────┴──────────────────────┘

Prompt caching: −90% on input tokens (all models)

US-only inference (Fable 5): +10% on input and output

Output is always 5× the input rate across all models

(I have not idea how to format this properly but the ASCII is fine)

This is interesting: there is a mountain of data (eg code in gh) that is the "truth" because God labeled it as so: meaning that many people are using it to do something (they have voted).

There is also a mountain of bullshit eg "[architectural|design] [anti]patterns" that are written mostly (I would argue) to sell something (consulting, hardware, etc). This is typically at odds with a good solution.

There is a relative lack of actual documented architectures that work. Not only do you need the details but also the usage of these systems so as to judge what "good" is.

We will probably just go the HTML route with architecture: take a really bad base and just keep throwing compute, memory, and network I/O at the problem until it works.

Note to self: invest in energy ETFs.

Or in a more general sense: Get your code as close as you can to your data

So says Gemma 3.