tripplyons

You only need to cache for tokens that are actually being used. Using 100k tokens of a 256k token window takes the amount of memory as using 100k tokens of a 1m token window.

For OpenAI, you have to use the Enterprise plan at API pricing in order for them not to train on your data.

Source: https://chatgpt.com/codex/pricing/?type=team

I use a $200/mo OpenAI Codex sub. Throughout the workday I run an average of 2 concurrent agents of GPT 5.5 with high reasoning on fast mode and use less than half of my subscription usage.

For more interactive/active usage you might be better off using the low reasoning level, but I have usually found high to be a good balance of intelligence and generation speed.

For context, here is the extension it made me: https://gist.github.com/tripplyons/ec953181707b6813d4be9e934...

I like to use Pi (https://pi.dev/), and I recently got it to make an approval extension for itself. It has a lot of documentation built-in for the agent to modify the behavior of the app.

I got it to display all proposed file change diffs and bash commands and made it so I can either approve the action or deny it with a message for it.

It was surprisingly easy to tell it to modify things things the diff viewing algorithm or syntax highlighting for the diffs.

Yes, this has been a highly requested feature! This will be the next feature for me to add after addressing the bug reports.

I have not encountered this, but I will note it down as something to look into. Which potion was it?

Thanks for the info! I will collect all these reports today and get these issues fixed in the next few days when I can.

Thanks for the info, I will look into this! I got past that battle and further into the game in my testing, but I don't think I've tried using an item in battle yet.

I chose Pokemon Emerald because it is my favorite of the games that have been disassembled!

Yeah, I made sure saving worked correctly

Yes, this project was made in around 15 hours of Codex.

Yes, it a recompilation of a community decompilation!

I have not added that yet, but it would probably be quite easy to throw a few prompts to Codex to do so.

The reduction rules seem kind of arbitrary to me. At that point why don't you just use combinators instead of defining a set of 5 ways their operator can be used?

I'm honestly surprised the CFO of Oracle doesn't make more than $950K

Your submission history looks like a bot trying to get engagement or something.

I think it's more about resisting some humans than it is about resisting machines.

I saw a video of guy who became an Amazon bestseller in a book category pretty easily by buying his own book.

MLA makes it so the keys and values used are a function of a smaller latent vector you cache instead of a key and a value for each token. KV cache quantization reduces the size of the values in the cache by using less bits to store each value. These two approaches operate on different parts of the process so they can be used in combination. For example, you can quantize the latents that are stored for MLA.

There are papers that try to quantize angles associated with weights because angles have a more uniform distribution. I haven't read this specific paper, but it looks like it uses a similar trick at a glance.

The good thing is you can still use their software without signing in and having to agree.

Got it.

The scenario I'm describing there is how a service like AWS has the ability to tamper with your code or its output. If instead, each response came with a ZK proof showing that the inputs you provided lead to the outputs it returned, you could efficiently verify that nothing was modified.

As much as I like the ideas, this article looks AI generated. This line with the bullet point, bolded label and colon, em-dash, and the second clause "it's about" all point to AI writing.

"Fiber-optic cables: Fiber-optic cables enable higher bandwidth phone lines and television—it’s about getting more television channels to more people."

One problem with private age verification is that because each verification cannot be traced back to a user, it is hard to prevent abuse like credential sharing. Imagine how a single stolen credential can be used by any number of users because the verification step kept the credential private.

To answer your question, ZKPs can enable the verification step to be done privately in your example. Another use case could be allowing cloud computing hosts to prove that they did not tamper with the results of a computation.

Saying Jane Street caused a $40 billion loss is wrong. Terra caused the loss because it was a Ponzi scheme that claimed to offer 20% APY on a stablecoin that wasn't backed by any real dollars.

It's a bad headline. They used publicly available blockchain transactions and didn't cause the collapse of the Terra ecosystem. Terra collapsed because it was a Ponzi scheme offering 20% APY on a fake stablecoin. The Terra stablecoin was not backed by real dollars, but instead by a cryptocurrency called Luna that did nothing else other than let you issue Terra stablecoins.

ZKML is a very exciting emerging field, but the math is no where near efficient enough to prove an inference result for an LLM yet. They are probably just trying to sell their crypto token.

It is definitely true across different chips. The best kernel to use will vary with what chip it is running on, which often implies that the underlying operations will be executed in a different order. For example, with floating point addition, adding up the same values in a different order can return a different result because floating point addition is not associative due to rounding.