dpflan

At this point, you may as well get a powerpack for a mini and put it in one of these slings, you could have a crazy powerful machine in your "sock-et" sling thing here...

When the iPhone Air was just another huge phone...but thinner...smh. Apple should put up some page to check interest level in a smaller phone, and with enough interest, go manufacture it. If it is more expensive because economies of scale don't work out, but they create one that is small yet powerful, that's what I would buy at premium, because apparently compactness is a luxury.

Yes, applied research has yielded the modern expert system, which is really useful to experts who know what they are doing.

Thanks for the details. Also, always appreciated Discord's engineering blog posts. Lots of interesting stories, and nice to see a company discuss using Elixir at scale.

What are you using it for? Is it part of a hybrid search system (keyword + vector)?

You are model citizen with those file moves. Totally agree with how that can be very disruptive for legibility and comprehension of changes.

`git-bisect` is legit if you have to do the history archaeological digging. Though, there is the open question of how git commit history is maintained, the squash-and-merge vs. just retain all history. With squash-and-merge you're looking at the merged pull-request versus with full history you can find the true code-level inflection point.

Agree, better title for this post; the fact that back prop is a leaky abstraction is a reason one should understand it and know how to do the mechanics by hand to truly experience it and develop understanding and intuition. Software / code abstracting away even more of the process leaves it open to magical thinking. I had to do hand calculations and convolutions in my Deep Learning graduate school course.

Thanks for the clarifying response.

Sounds like The Paradox of Choice.

- https://en.wikipedia.org/wiki/The_Paradox_of_Choice

Are you able to provide more context and details? It’s always interesting to learn about system design and architecture decisions from real world examples.

Thanks for sharing. Yeah, I think that's a problem with bolted on functionality. Did you end up building a new version with a different language? What kind of problems did you encounter that made the solution "not great"?

> It's very natural and ergonomic in Elixir, because it's the paradigm of choice in the underlying vm.

Thank you, this is the key factor -- a fundamental design decision that influences the foundation.

Correct, it's not strictly unique to this. Even Ruby has it, but Ruby does not have BEAM/OTP.

Smalltalk was a fun language to work with, albeit academically.

When you consider the message passing paradigm, I can envision how that simulates neuro-chemical communication between neurons (now here being functions that do things). So there is a design for communication between functions, then there are memory storage and information retrieval parts of the system, short-term RAM, long-term HD/SSD, databases, cache-systems, where relevant information can be and become manipulated. Long and short running processes, acceptance of fail-fast and that communications can fail and that's just part of software system life (I assume also a similar idea in the human brain, don't you find yourself forgetting something you were just thinking about?). There is then the external part of the system, accepting inputs from the outside.

Elixir is pretty "nifty", and has Rails feels. I have worked on and seen its performance compared to scaled up Rails application. The BEAM / OTP is a very cool system to program on top of. I think it would be a great underlying system to develop a consciousness for AI systems with its concurrency paradigms and message passing, node connecting, other builtins. I'm not sure if the AI/ML focused Elixir projects like Nx have really taken off, but again, an Elixir based system that accepts numerous connections, passes messages around like exciting synaptic communications between functions... it just seems cool, maybe just on paper...

The author is a self-professed front-end developer according to LinkedIn, which may be influencing experience with backend systems and functionality. I was wondering really why this terse post is sky-rocketing to #1 today, hopefully we get some good discussion from the community.

Interesting. What aspects of the development workflow/cycle have the most room for improvement (i.e. is there ranking of the "height" of the "hanging fruit" throughout the process)? What sort of software tooling would help?

True, from scratch, may need to be cautious regarding re-creating it having built it on your employer's time and machinery/property, check your employment contract. Still very cool that you built this!

Do you have paying customers? Putting money into customer acquisition and retention is where a lot of money can go versus R&D.

Is there a visualization possible of this pattern?

For some reason this made me think of the Ulam Spiral -- https://en.wikipedia.org/wiki/Ulam_spiral.

Yes, AI allows for exquisite demos, demos that tantalize the audience into thinking of the infinite potential of the technology, that stunning vision expands and expands until the universe of potential overwhelms the dreamer into a state of terminal fantasy. So it is always a solution looking for a problem. There are cases where the two meet more realistically and a valuable impactful company develops it.

The fact it can generate human language that is very compelling for certain context, makes it seem possible of doing so for many, many more contexts.

Thanks for sharing, any particular articles that had last impact on you?

Has anyone read up on the recent paper from Meta/FAIR -- CWM: An Open-Weights LLM for Research on Code Generation with World Models

Which looks to attempt to give better "coding" understanding to the model instead of mere tokens and positioning and hence improve the coding capabilities of these "brilliant but unpredictable junior engineer" coding agents:

- https://ai.meta.com/research/publications/cwm-an-open-weight...

Sometimes I wonder if pseudocode could be better for prompting than expressive human language, because it can follow a structure and be expressive but constrained -- have you seen research on this and whether this an effective technique?

> Most of my initial prompts to agents start with "DO NOT WRITE ANY CODE YET."

I like asking for the plan of action first, what does it think to do before actually do any edits/file touching.

I think this would benefit from examples of including coding assistants in the stages enumerated; how can the agent be included in each stage? I've seen posts about successful collaboration with agents at say Google, where there is tons of upfront work among humans to agree on design, then work with the agent to build out parts of the project and ensuring thorough test suites are included.

Does including an agent at each stage of this cycle mean "context engineering"? Is this then just more text and assets to feed in at each stage of LLM ussage to provide the context for the next set of tokens to generate for the next stage of the cycle? Is there something deeper that can be done to encode this level of staged development into the agent's weights/"understanding"? Is there an established process for this yet?

- Specification

- Documentation

- Modular Design

- Test-Driven Development

- Coding Standard

- Monitoring & Introspection

sshhh ... don't sidetrack the productive comment generation. (also, SSHHH as a possible name)...

I recall the noted Zuckerberg comments regarding the situation you describe of why people are willing to trust you with their privacy and data...

Or send a prompt injected spam in someone's GMail, doesn't even have to opened by the human end-user:

https://www.linkedin.com/posts/eito-miyamura-157305121_we-go...

Here is the WSJ opinion piece: https://www.wsj.com/opinion/is-gen-z-unemployable-288d2ec9

Here is a LinkedIn post by one of the researchers, Suzy Welch at NYU: https://www.linkedin.com/posts/suzywelch_opinion-is-gen-z-un...

Pair these results of this research with the impact of AI on entry-level jobs, and it does certainly questions about the future.

Also, add in this observational piece by Hunter S. Thompson, written after 9/11 for more color about Gen Z's world: https://www.espn.com/page2/s/thompson/010918.html

"""

We are At War now, according to President Bush, and I take him at his word. He also says this War might last for "a very long time."

Generals and military scholars will tell you that eight or 10 years is actually not such a long time in the span of human history -- which is no doubt true -- but history also tells us that 10 years of martial law and a war-time economy are going to feel like a Lifetime to people who are in their twenties today. The poor bastards of what will forever be known as Generation Z are doomed to be the first generation of Americans who will grow up with a lower standard of living than their parents enjoyed.

That is extremely heavy news, and it will take a while for it to sink in. The 22 babies born in New York City while the World Trade Center burned will never know what they missed. The last half of the 20th century will seem like a wild party for rich kids, compared to what's coming now. The party's over, folks. The time has come for loyal Americans to Sacrifice. ... Sacrifice. ... Sacrifice. That is the new buzz-word in Washington. But what it means is not entirely clear.

"""

- HST