I have vLLM running on a Linux machine in my basement, connected with Tailscale, and I use small models as part of tasks like this:
- Transcribing scanned documents into formatted text
- Captioning/describing images and classifying them for audience suitability (includes anti-spam)
- Matching documents with relevant Wikipedia pages for tagging
I don't use them like frontier models. I break the work down into micro-tasks with one clear goal for each prompt. I write a lot of glue software to make the complete flow work. I was working on all of these tasks before LLMs appeared on the scene. The LLMs have allowed me to replace a lot of complicated code with less code plus a model, while achieving better results.
I use local models for reasons of cost and control. I already had the workstation and GPU. The only running cost is electricity. I have used proprietary models from OpenAI and Google for some of these tasks, but I also encountered churn when the models I built my tools around were retired. I don't worry about that when I have the weights saved locally.
The window handles, on the other hand .. this was correct in Windows 3.0 and there's basically no good reason to have changed it. There should be a title bar. Active window should have visibly contrasting title bar. There should be sufficient grab space all round a window to get hold of it.
Bonus points: move your mouse pointer very slowly around a bottom curved corner window handle on Windows 11. Ask yourself: how well does "place I am pointing at" line up with "where the curve is"?
The RK3576 is a really interesting/versatile chip and it is awesome to see major effort going into baking full support into the linux kernel. I could see it opening up a ton of doors for awesome FOSS hardware projects w/ AI accelerated workloads.
One idea I have (but realistically will probably never build) is an e-ink notepad with a microphone that I can ask to generate custom note-taking templates. As a niche example, I'm imagining I'm at a baseball game and I can tell my tablet "hey give me a baseball scorecard template" only for it to generate one for me. Then if there are a ton of subs or the game goes super long, I can modify the template in place with follow-up commands like "add more rows for player substitutions" or "make it support up to twelve innings".
I imagine having a chip like the RK3576 fully supported in the linux kernel could make building a device like this much much easier.
This page is copyright 2005 by Graeme Cole. What are you allowed to do with it? Pfft. Anything within the realms of common sense, really. I don't want to prescribe rigidly what people can and can't do with it, so I've decided on a benchmark. It's this: you're allowed to do with this page anything you wouldn't mind me doing with your cat. So yes, you can photoshop it for comedy effect, you can copy bits of it for illustrative purposes and so on, but you can't steal it and pass it off as your own.
The "contrarian dynamic" is that HN threads (and internet comments generally) are propelled by people making objections. [1]
The objections come in waves. In the earliest stage of a thread, they tend to be rapid negative reactions to the article. It's not that these are a community consensus, it's that they're the fastest reactions to feel, and the fastest comments to write—especially when the topic is provocative, when most of us are reacting from cache [2].
Then a second wave of objections is generated by the first wave.
Readers come to the thread, see the comment section dominated by those initial reflexive responses, and feel some version of surprised-shocked-dismayed at how "all HN commenters" seem to react in the same way. This propels them to write defenses of the article, often carefully expressing more moderate or balanced views than the first wave—but they probably wouldn't have been motivated to post anything if there hadn't been the first wave of comments to object to!
These second-wave comments tend to get more upvotes, perhaps because more people tend to share the more moderate view, but also because those comments tend to be more reflective [2] and therefore better written.
This explains why the top comment in a thread so often begins (ironically) with "Wow, I can't believe the comments here"—or from the current thread: "All of these comments make me think HN has never interacted with a 5 year old" [3]—followed by a defense of whatever those objections were objecting to.
Eventually you get objections to the objections to the objections—which reminds me of the line "My complication had a little complication" from Brazil [4], and also epicycles.
Other people have recommended getting into support, and it really is a good option to consider.
A lot of the suggestions here are a bit more long term. There are entry level support jobs in tech that will give you valuable experience, _now_, for a pathway to engineering roles. Many companies will also give you a budget to buy courses, books, conference tickets, etc. Staying in support isn't a dead end, anymore, either. Remote companies will try to pay you based on location, but salaries can get very high once you stack up experience (120k+/year).
Regardless of future plans, you'll have an advantage over trying to learn while doing unrelated work.
1) Collect job postings for jobs you are interested in.
2) Look at the "skills" portion of the job postings.
3) Assess your level of proficiency in those skills.
4) Practice each skill until you feel confident you have a working level of proficiency in those skills.
Part II: Study tips
1) Take note of every acronym or concept you don't understand. Then, look it up and try to come up with your own definition for it. Write down that definition in your own words. At first, that definition may not be correct, but you have to try to keep it as accurate as you can.
2) When you study a concept from a book or website, also take notes. Then, write what you have learned about each topic. Force yourself to explain what you are doing and why.
3) Vocalize what you are doing and what you are thinking. Record yourself and then listen to yourself. Expressing your thoughts in a coherent way will help you later during interviews.
4) Enroll into a MOOC for Algorithms and Data structures.
Part II: Resume
1) Create a 1 page resume that emphasizes your skills and abilities.
2) In the absence of formal education, create a portfolio that showcases your abilities.
3) If you have completed a MOOC, you can list it there.
Part III: Apply
1) Look for referrals among people you know.
2) Announce you are looking for a job.
3) Apply and interview.
Part IV: Interview
1) Don't get too nervous. Stress management is part of the job.
2) Budget your time. Questions may have follow up questions. Ask if there are follow up questions in advance.
3) Explain your thoughts clearly.
4) Ask questions, the interviewer can offer help if needed.
5) Learn about the company, what they do, thier mision, vision, core values, etc. Crunchbase can help you find basic information about a company.
4) Prepare to ask questions. e.g.: what is their engineering organizational structure, what is their culture like, etc. Prepare to answer questions such as why do you want that job, why do you want to work there, etc.
Apply to a lot of them, there is a statistical element to this. Apply to smaller, non-tech companies where competition is not as high. Make your resume as good as possible while being honest.
This may be overly personal, but I would consider contract/temp work, too, to build up your professional experience. That would also reduce the risk of losing your existing full-time job and ending up in a financial hardship.
Also, apply to non-developer roles. There is a bit of complexity to working in a software company, more than just writing code. Consider a support or other development adjacent field that is still technical.
Hey anon, I feel for you, and most of the answers here are giving you a 12-month plan when you need a 1-month plan.
If you know Linux well enough to configure Wordpress, how about tech support for a hosting company? Try making some short email approaches to old-school hosting companies (I used to run one). They might be flexible enough to take a chance depending on how confident your approach is.
They have lots of customers bashing away at Linux, making mistakes, often without the patience to see their own problems through. Their business problem is that these customers need hand-holding but only pay a fixed, monthly fee. (the hope is eventually they stop asking and keep paying for years).
The combination you can offer those companies could be basic Linux knowledge (no need for advanced cloud stuff) and whatever flexibility you can offer them - especially if they're not in your time zone.
The larger ones might be a tall order, all listing locations by default (Gandi, Leaseweb, GoDaddy, Hetzner etc.), but maybe someone here will have an inside track.
I agree with other posters - never mention difficult circumstances in a job application, particularly a cold application. Just talk about how keen you are to solve their difficult customers' problems, how well you work in a team, and keep the initial approach brief.
Nice! I had an idea that I never coded up. Basically a hamster with a parachute that would fall slowly to the bottom of the page. If you scroll down fast his parachute will deploy to catch up and land once again at the bottom of the viewport. He'll hold a "Top" sign when you hover over him and will shoot off with a jetpack when you click on him and as the screen hits the top you'll see him deploy a parachute again to gently land at the bottom of the screen.
Wonderful! What a cool idea. For anyone interested, you can learn the whole of Hangul in an afternoon; it's cleverly designed to be very logical and has some handy mnemonics: https://korean.stackexchange.com/a/213
Fun story - at Oxford they like to name buildings after important people. Dr Hoare was nominated to have a house named after him. This presented the university with a dilemma of having a literal `Hoare house` (pronounced whore).
I can't remember what Oxford did to resolve this, but I think they settled on `C.A.R. Hoare Residence`.
I've tried getting started with this but my first attempt a habanero/mango sauce was _horrible_, must've used a slop recipe or something. Do you have a good base to recommend?
1. In the immediate aftermath, news outlets will get it wrong.
2. Don't trust anonymous sources.
3. Don't trust stories that cite another news outlet as the source of the information.
4. There's almost never a second shooter.
5. Pay attention to the language the media uses.
• “We are getting reports”… could mean anything.
• “We are seeking confirmation”… means they don’t have it.
• “[News outlet] has learned”… means it has a scoop or is going out on limb.
6. Look for news outlets close to the incident.
7. Compare multiple sources.
8. Big news brings out the fakers. And photoshoppers.
9. Beware reflexive retweeting. Some of this is on you.
Arm's biggest customers are also their biggest competitors. Apple and Qualcomm use their ISA license instead of their Core architecture license. ISA license pays a much smaller royalty fee.
Now that Apple and Qualcomm both design their own custom cores, Arm's custom core license business is suffering. The only other company that license their custom cores for phones is Mediatek. Arm still has hyperscalers in Google, AWS, Microsoft, and Meta for server core licenses though. But that might not last as long since everyone wants to differentiate and design their own cores. Apple and Qualcomm are succeeding in designing better custom cores than original Arm cores.
So Arm's business model has to change if they want to survive.
Either they make ISA license fees much higher (risk losing customers to RISCV) or they make their own CPUs and compete against their own customers even more.
I enjoy analyzing chip companies and businesses. I've never invested in Arm's stock because their customers are also their competitors. Ultimately, the questions for Arm investors are:
1. Do you believe they can design better cores than Apple and Qualcomm (and other big tech)?
2. Do you think that the threat of companies moving to RISC-V is real if Arm raises ISA royalties?
Once men turned their thinking over to machines
in the hope that this would set them free.
But that only permitted other men with machines
to enslave them.
...
Thou shalt not make a machine in the
likeness of a human mind.
-- Frank Herbert, Dune
You won't read, except the output of your LLM.
You won't write, except prompts for your LLM. Why write code or prose when the machine can write it for you?
You won't think or analyze or understand. The LLM will do that.
This is the end of your humanity. Ultimately, the end of our species.
Currently the Poison Fountain (an anti-AI weapon, see https://news.ycombinator.com/item?id=46926439) feeds 2 gigabytes of high-quality poison (free to generate, expensive to detect) into web crawlers each day. Our goal is a terabyte of poison per day by December 2026.
Join us, or better yet: deploy weapons of your own design.
> Do not fall into the trap of anthropomorphising Larry Ellison. You need to think of Larry Ellison the way you think of a lawnmower. You don't anthropomorphize your lawnmower, the lawnmower just mows the lawn, you stick your hand in there and it'll chop it off, the end. You don't think 'oh, the lawnmower hates me' -- lawnmower doesn't give a shit about you, lawnmower can't hate you. Don't anthropomorphize the lawnmower. Don't fall into that trap about Oracle. — Brian Cantrill (https://youtu.be/-zRN7XLCRhc?t=33m1s)
And
> I actually think that it does a dis-service to not go to Nazi allegory because if I don't use Nazi allegory when referring to Oracle there's some critical understanding that I have left on the table […] in fact as I have said before I emphatically believe that if you have to explain the Nazis to someone who had never heard of World War 2 but was an Oracle customer there's a very good chance that you would explain the Nazis in Oracle allegory. — also Brian Cantrill (https://www.youtube.com/watch?v=79fvDDPaIoY&t=24m)
You are doing exactly what happens repeatedly in these sorts of conversations. You completely ignore all subtext, all historical context, all conclusion, all metatextual reading in favor of exact out-of-context literal wordings which you claim will convince you of the case. That is not what will happen, but of course you know that already.
No xenophobic racist with an ounce of sense is going to write something like "brown people are rapist criminals who have invaded our lands and must be cleansed". It's absurd to maintain we must provide that as proof or it didn't happen.
I'm referring to the far right ramblings of DHH supporting political figures who are associated solely with violence and intolerance.
I'm referring to the far left activists who have developed a habit of coercing individuals, organizations and communities, sometimes with threats of physical escalation, for perceived connections to other political fringes.
I'm referring to those using language along the lines of "they have hitler particles in them".
I'm referring to tech journalists who do nothing except whine about left wing politics, however benign.
It's mind melting.
---
Edit: To add colour to the final reference, the same journalist who attacked Framework for issuing Pride stickers is now coming to their defence. It's predictable, it's unproductive, and it should be filtered by anyone who values their time.
That would be very dangerous and possibly the plot of some hard sci-fi thriller. If immortal mice were to somehow escape the lab and breed in the wild we could rapidly end up with massive populations of mice that never die of aging, ultimately collapsing entire ecosystems and food chains, leading to the end of humanity.
- Transcribing scanned documents into formatted text
- Captioning/describing images and classifying them for audience suitability (includes anti-spam)
- Matching documents with relevant Wikipedia pages for tagging
I don't use them like frontier models. I break the work down into micro-tasks with one clear goal for each prompt. I write a lot of glue software to make the complete flow work. I was working on all of these tasks before LLMs appeared on the scene. The LLMs have allowed me to replace a lot of complicated code with less code plus a model, while achieving better results.
I use local models for reasons of cost and control. I already had the workstation and GPU. The only running cost is electricity. I have used proprietary models from OpenAI and Google for some of these tasks, but I also encountered churn when the models I built my tools around were retired. I don't worry about that when I have the weights saved locally.