rileyphone

DeepSWE is closer to that

https://deepswe.datacurve.ai/

Absolutely, plus if you control the coding agent you can enforce certain guarantees and have it wrap your services with a custom sdk. I've been exploring this pattern in a couple of different domains where it's just a vite react app wrapped in an iframe with a JWT bridge giving auth, hosted on a separate domain.

Here's one: AI democratizes the ability to produce software, which has mostly been an arcane craft wielded by a priestly class. Now anyone, if they know what they want and it isn't too complex, can talk to AI and get working (if not also janky) software in a very short amount of time. Hopefully this breaks the grip that platforms/large corporations have on personal software and the internet.

All that says is some benchmarks aren’t worth the tokens it takes to evaluate them. Mythos is clearly capable of finding zero days other models can’t, and Fable is close enough to be lumped with it.

Pretty neat! Really like the design and minimalism. Oh, and the free hosted models :-) (but I assume that's just for the demo). Kind of confused by the open in cursor button but otherwise clean.

It seems like there's a lot of stuff out there that's similar, but it's all either focused on art/roleplay on the one hand, or enterprise teams on the other. Most prompts I see are shared as raw text on twitter or something, to be lost in my bookmarks...

Anyways, here's a prompt. Next time I see an actually useful one out there I'll hopefully remember to save it on Prom.

https://prom.dev/p/where-is-mama

Yes but it's really good for building better object systems, especially with newer features like proxies. In the SELF days they implemented Smalltalk with prototypes and found that it was faster than normal Smalltalk implementations.

https://www.researchgate.net/profile/Mario-Wolczko/publicati...

According to this delightful overview [1] of the desalination plant, the capacity overall is 12000kW so that's definitely close enough.

1. https://www.niph.go.jp/soshiki/suido/pdf/h21JPUS/abstract/r9...

Location: Seattle, WA

Remote: Sure

Willing to relocate: No

Technologies: Javascript (Node/Web/Bun), Linux, Python, SQL (SQL Server/MySQL/Postgres/Sqlite), Azure, k8s/docker, React, Typescript, and many more

Resume: https://rileystew.art/files/resume.pdf

Email: me @ the above domain

Website: https://rileystew.art/

I'm a generalist with a focus on backend and infrastructure -- in my career that's meant a lot of diving into logs, probing systems, and solving distributed bugs. In my own time I am working on an object-oriented Javascript framework and a custom autoencoder for visualizing latent states of LLMs. Looking for the opportunity to build something great and solve interesting problems.

From talking to people, the average user relies on memories and chat history, which is not easy to migrate. I imagine that's the part of the strategy to keep people from hopping model providers.

I got a Sovol SV06 ACE a few months ago as it seemed to have most of the nice features of the Bambu (like auto bed leveling) without the closedness. The printer runs Klipper and you can ssh into it. So far there's been one issue where I had to replace a fan but otherwise it's been great. Much cheaper than a new Prusa too.

Page for the role doesn't have Seattle, only SF.

It's baked in to the process as part of design review, after getting pass the first wall of zoning.

https://www.seattle.gov/neighborhoods/public-participation/e...

Firefox has made Mozilla billions over its lifetime by selling the default search engine rights to Yahoo and Google. Chrome, having a much greater user base, would demand a correspondingly higher fee (probably around $10b a year). Now, the other problem is there is no other search engine to compete with Google at that level, but that might change with independence of Chrome.

If you spend 1000W on 1 gigaflops, but could have gotten 1 teraflops instead on newer hardware, you are mostly just throwing money away. Unless fab capacity is severely limited in the future or energy becomes too cheap to meter, the opportunity cost is just too great for your statement to be true.

Been working on an autoencoder that converts the hidden states of transformer models into a spatial representation that can be visualized. Started more on the toy scale but now I'm trying to scale it beyond my humble 3060. Using LLMs to help with torch and such but they are limited in the details of tensor twiddling.

https://github.com/ristew/weightscan

Gemini is probably using ring attention. But scaling to that size requires more engineering effort in terms of interlink that goes beyond the purpose of this release from Mistral.

According to the paper, the dataset goes up to 1978 because that's when copyright law was updated to automatically apply to newswires. It's unfortunate that we got into the situation where academia has to play by the rules wrt copyright while big private labs flaunt it.

In that case there are two attractors - one towards the Golden Gate Bridge and one towards the harmless, helpful, honest assistant persona. Techniques as such probably get weirder results with model scale but no reason to think they get wiped out.

In this case Azure is responsible for the datacenters, billing, and support.

There's also the oceanic currents carrying warm water from the gulf. Europe is in a very lucky position.

Unary operators are still operators. The integer parsing rules are probably different. In Lisp, -x would be a symbol, and the proper analog to -123 would be (- x) eg.

Not really because the ordering is unambiguous given the parens. No need for operator precedence rules like PEMDAS.

The bigger size is probably from the bigger vocabulary in the tokenizer. But most people are running this model quantized at least to 8 bits, and still reasonably down to 3-4 bpw.

To be fair, the Llama 2 instruction tuning was notably bad.

$10 a month?

128GB in one chip seems important with the rise of sparse architectures like MoE. Hopefully these are competitive with Nvidia's offerings, though in the end they will be competing for the same fab space as Nvidia if I'm not mistaken.

With an MoE you only need to train a smaller model which you can then combine into an x8 and finetune/train the router. Mistral used their 7B base to make Mixtral, Qwen's new MoE uses their 1.8B model upscaled to 2.7B, pretty sure Grok also trained a smaller model first.

LLMs aren't nearly as monetizable (now) as search ads. Very possible that search collapses in the near future alongside the open web, and that whole line of revenue destroyed.

It's just such a bad name though.

Nice! I started a similar project with a rotary phone a couple years ago with asterisk (just to the point of making a test call). Wish I had heard about the Windstream device, I ended up getting a pulse to tone converter and a PAP2T to voipify it, but those seemed to work. If I decide to pick it back up again I will check out your scripts.