One is a TRS-80 Model I emulator in JavaScript called Trash80. About 10 months ago I started this project just for fun while experimenting with what now seems to be called agentic loops. I got things working pretty well with the Z80 passing the ZEXALL suite and a lot of real TRS-80 software running fine. It sat for months untouched before I decided it is worth releasing and recently started it up again.

I didn't want to release it without a ROM, so I rigged up some agents to build a clean-room style L2 ROM w/ a fairly complete BASIC and even readline-style control commands, history, and a proper cursor. That went very well, but the agents cheated on floating point and implemented some weird Q5.2 like-thing. I told them to fix it, but I guess I didn't give clear enough instructions because they replaced it with a BCD hybrid monstrosity instead of proper floating point. The proper floating point is now underway, but I'm mostly using excess Codex credits before they expire, so it's only moving forward when I have credits I don't need.

I also built a silly ASCII fractal browser in Z80 assembly so that I can ship with a virtual disk that has software on it. The emulator works in the browser and the terminal. Unicode sextant block graphics map very well to TRS-80 Model I semigraphiccs/squots, so it really does run everything very well in the terminal, even games. I also added a line-mode for line-based applications, so you can use a readline-like interface and feel like it's native terminal app as well, though that has some issues I need to fix. And of course, you can shebang TRS-80 BASIC files and run them through the emulator too.

Another project was a demo of chromesthesia, a form of synesthesia where sounds trigger experiences of color. I thought it was done and ready to release, but then I had a new idea. The visualization while cool, was kind of boring. I decided to replace it with an attempt at a semi-physically accurate cymatics simulation with artificial coloring based on chromesthesia. Cymatics is the practice of making sounds visible by vibrating a surface, such as a plate with sand on it. As the sound changes, symmetrically interesting patterns form and evolve. I've got something working now with wave generation and microphone input, but sometimes it gets a bit stuck and stops evolving as it should, so I have to find time to figure that out.

Currently all unreleased, but when they do release it will be at www.leshylabs.com. I sometimes post updates on X, but not too often. (https://x.com/LeshyLabs)

This predates the WebAudio API, and so it builds data URLs as WAVs instead. Every sound can be represented as a short string, and the tool can be used as a library that procedurally generates the sounds from the strings. It also has a "song mode" where notes can be provided to guide the sound, which makes certain types of more complicated sounds possible to make, including jingles and short songs.

It's definitely dated now, but if anyone is interested, you can find it here

https://www.leshylabs.com/apps/sfMaker/

See the "Example Sounds" at the bottom of the page to hear what it can do.

I made some AI tools (https://github.com/DougHaber/lair) and added in a tmux tool so that LLMs could interact with terminals. First, I tried Nethack. As expected, it's not good at understanding text "screenshots" and failed miserably.

https://x.com/LeshyLabs/status/1895842345376944454

After that I tried a bunch of the "bsdgames" text games.

Here is a video of it playing a few minutes of Colossal Cave Adventure:

https://www.youtube.com/watch?v=7BMxkWUON70

With this, it could play, but not very well. It gets confused a lot. I was using gpt-4o-mini. Smaller models I could run at home work much worse. It would be interesting to try one of the bigger state of the art models to see how much it helps.

To give it an easier one I also had it hunt the Wumpus:

https://x.com/LeshyLabs/status/1896443294005317701

I didn't try improving this much, so there might be some low hanging fruit even in providing better instructions and tuning what is sent to the LLM. For these, I was hoping I could just hand it a terminal with a game in it and have it play decently. We'll probably get there, but so far it's not that simple.

Lair currently consists of three main components.

First, there is "chat", a command line chat interface with autocomplete, multi-line editing, markdown rendering, persistent sessions, etc. It's really more than just a chat interface, and I use it for prompt engineering and testing things out against different models quickly. This is built using the Python PromptToolkit and Rich modules, which I can't recommend enough. Most things are configurable, and there are lots of keyboard shortcuts available.

Next up we have "comfy". This uses ComfyScript (https://github.com/Chaoses-Ib/ComfyScript) to run workflows against a ComfyUI server. It can do image and video diffusion as well as upscaling, with a CLI that works well for scripting. ComfyScript is great, but does have some rough edges that might need to be smoothed out to get better performance, easier debugging of errors, and more control over what it writes to STDOUT/STDERR. Before I made an open source version of Lair, I was implementing a lot of this functionality by hand. It's been really nice letting ComfyUI provide features and optimizations so that I can worry more about providing an interface on top of it.

Finally, there is "util". This is a one-shot tool for LLM interactions from the command line and for shell scripting will LLMs. I've seen a number of similar tools elsewhere. This one is in the style I like, and it works with most Lair features, so it supports attachments (PDF, text, and image with vision models,) sessions, and all of Lair's supported tools.

The README.md has lots of examples of the functionality and there is also a Youtube video in the Overview section with a live walk-through. I hope someone else finds this useful. I am happy to answer any questions.

Comments URL: https://news.ycombinator.com/item?id=43399442

Points: 2

# Comments: 0

We are building a best-in-class advertising platform for brands and advertisers to take full advantage of the massive reach on television. This is an opportunity to join a team of amazing engineers and data scientists.

We are currently looking for:

* Senior Software Engineer, Core Platform

https://jobs.lever.co/simulmedia/3f5416b2-b43b-4145-8017-fbb...

* Senior Software Engineer, Applications

https://jobs.lever.co/simulmedia/9c18cea4-d874-44c1-9a9f-343...

* Senior Software Engineer, Data Services

https://jobs.lever.co/simulmedia/25aa75bf-ee70-4ad4-8bf2-726...

* Senior Software Engineer, Machine Learning

https://jobs.lever.co/simulmedia/a0337745-ac19-4b59-b3e1-c84...

Please apply directly through: https://www.simulmedia.com/careers/

For those that don't know, the Brave browser has Tor tabs, which route through Tor. It also has the standard private tabs. Tor support currently exists only on the desktop Brave browser.

Here is the announcement: https://brave.com/tor-tabs-beta

Brave has been supporting Tor, and running Tor relays to improve the network.

Brave is newer at the game. They have had Tor tabs less than a year. They can do fingerprinting protection and no-script, but it's still a full featured web browser, with a lot of risks. The fingerprinting protection isn't as good as the Tor Browser, and unless they changed something, Javascript wasn't disabled by default in Tor tabs.

The Tor Browser has been around for a while and is meant to be a secure web browser from top to bottom. It has had a lot of development looking to find and fix possible leaks and to ensure security. That is its primary focus, and it is pretty good at it.

If you want to use Tor casually, maybe access an onion site, or just get a big boost in your level of privacy, the Tor tabs in Brave are a nice option. They are really easy to use and give great privacy. It is good for casual Tor use.

If you want (or need) serious privacy, the Tor Browser is a better choice. That is its purpose. It is developed to be hardened for protecting the user and it will provide better protection.

Our tech stack is written in Ruby, Python, and Javascript. We deploy microservices in Docker to AWS.

We are currently looking for:

* Software Engineer, Infrastructure (applications and tools development on an SRE / DevOps infrastructure team)

* Software Engineer, Front End - Applications

Please apply directly through: https://www.simulmedia.com/careers/

We are building a best-in-class advertising platform for brands and advertisers to take full advantage of the massive reach on television. This is an opportunity to join a team of amazing engineers and data scientists.

Our tech stack is written in Ruby, Python, and Javascript. We deploy microservices in Docker to AWS.

We are currently looking for:

* Software Engineer, Infrastructure (applications and tools development on an SRE / DevOps infrastructure team)

* Software Engineer, Applications / Full Stack

Please apply directly through: https://www.simulmedia.com/careers/

We are building a best-in-class advertising platform for brands and advertisers to take full advantage of the massive reach on television. This is an opportunity to join a team of amazing engineers and data scientists.

Our tech stack is written in Ruby, Python, and Javascript. We deploy microservices in Docker to AWS.

We are currently looking for:

* Software Engineer, Applications / Full Stack

* Software Engineer, Infrastructure (applications and tools development on an SRE / DevOps infrastructure team)

* Data Scientist (ML Pipeline / Operations Research)

Please apply directly through: https://www.simulmedia.com/careers/

After the change was announced, a number of people said they would fork and maintain the LGPL versions. One of the people behind EXT JS showed up in online discussions at the time and insisted that would be a violation.

The problem came from it not really being under LGPL. They tacked on this extra piece:

  Ext is also licensed under the terms of the Open Source LGPL 3.0 license. You may use our open source license if you:

   * Want to use Ext in an open source project that precludes using non-open source software

   * Plan to use Ext in a personal, educational or non-profit manner

   * Are using Ext in a commercial application that is not a software development library or toolkit, you will
     meet LGPL requirements and you do not wish to support the project

The EXT JS company in online forums insisted they were wrong, and further outraged the community. A lot of people, myself included, decided to stop using EXT JS. We were planning on the commercial license, but the response of the company didn't feel right, and so like many others, we abandoned EXT JS.

  Udo of Aachen (c.1200–1270) is a fictional monk, a 
  creation of British technical writer
  Ray Girvan, who introduced him in an April Fool's hoax
  article in 1999. According to the article, Udo was an
  illustrator and theologian who discovered the Mandelbrot
  set some 700 years before Benoît Mandelbrot.

  Additional details of the hoax include the rediscovery of
  Udo's works by the also-fictional Bob Schipke, a Harvard
  mathematician, who supposedly saw a picture of the
  Mandelbrot set in an illumination for a 13th-century carol.
  Girvan also attributed Udo as a mystic and poet whose
  poetry was set to music by Carl Orff with the haunting O
  Fortuna in Carmina Burana.

The algorithm is very simple. Each screen is treated as a separate level. Commands are randomly chosen, and anything that helps the agent move forward is saved. If the score goes down or the player's height goes below the surface, the game resets and a bit of history is trimmed.

Here is a youtube video of a trained path for 20 minutes: https://www.youtube.com/watch?v=I4j8xWq1Jsc

This will not play an optimal game, which requires using the underground and not accidentally jumping over treasure.

Comments URL: https://news.ycombinator.com/item?id=12164508

Points: 2

# Comments: 1

Through the eyes of a person compelled to give testimony, we may reach a different conclusion. Consider a journalist who is granted immunity and asked to give up their source. They refuse and spend six months in jail, with few to no options to appeal for their freedom. This isn't theoretical. It happens occasionally.

The definition of due process can vary from the common definition of fair treatment under the law, to the view more often taken by lawyers and governments that it just means following the process of the law as accepted by the courts and respecting legal rights.

The reporter probably would feel as though they weren't treated fairly under the law, but the prosecutor would argue that they followed the law in full. I probably should not have used the phrase "without due process" in this case, since that will make it difficult to communicate with anyone holding views of the government or lawyers.

They say their generous worst case calculation of losses is $550 billion, based off of the number of angel investors, throwing out a 1 million number against each, and then valuing companies at $327 billion.

They then compare that to the losses in NASDAQ of 5 trillion. That is based on the peak of 6.7 to the low of 1.6.

Why does the chart in the article only show a peak of about 5 trillion, which would cut the losses from about 5 trillion to to about 3.4 trillion?

Isn't it wrong to compare the estimate of angel investors and valuations against NASDAQ? Why not compare NASDAQ against NASDAQ? To use their style of comparing against extremes, back in 2009 we hit a low of 1.29. In July we hit a high of about 5.2. If those gains were erased, wouldn't we take a loss of about 3.91 trillion, which is a little larger than their chart showed for the dotcom crash?

Am I missing something, or is this article making a misleading comparison?

EDIT: Reading about the NASDAQ composite index on Wikipedia, it says that the index was changed in 2014, and the composition is very different then it was in 2000. So, even NASDAQ to NASDAQ may not work as a comparison, unless it can be recalculated the same way, and even then, other factors such as inflation should probably be factored in, since it was substantial. (The CPI shows $1 from 2000 is worth $1.39 in 2015 dollars.)

EDIT #2: After researching a bit, I think the difference between their chart showing a loss of 3.91 and them saying 5 might be because the 5 trillion is in terms of market cap. On August 18th we were at 9.15 billion, but I don't have a good source for historical market cap data. From the look of things, I'm guessing the losses off our recent peaks to the 2009 lows would be larger than the dotcom losses, but I'd love to hear from anyone who better understands these things.

I'm sure that some vendors will provide an option, so this will just likely be an extra thing to research before buying a new system.

I am looking forward to the future of Perl 6. I really do like some of the things I've seen going on there, and it is fairly likely that in the future I will be happily writing Perl 6 code on some projects. With that said, I am very concerned about how Perl 6 carrying the Perl name will impact Perl 5, but that is separate from my excitement that a new more modern language will be carrying forward some of the Perl styles.

In short, I do appreciate Perl 6 and I don't mean to downplay anything going on there, but my concern is about what it all means for Perl 5.