Then go with Qwen3.6 35B A3B. It's way faster (up to 5x) and it is 80% as capable as the 27B. The 27B is for serious people looking for one shot coding. The 35B is for iterative and quicker coding. I am in the same situation as you (making something I don't want to do myself) and I use the 35B at Q5_K_M.
Underrated habit - Even for people without diagnosed ADHD, there's a skill aspect to deep focus that is developed through experience and discipline. Habit is a shortcut to discipline.
What a lot of people see in themselves as ADHD is at least compounded by underdeveloped and (willfully) neglected skill in being able to focus. Not saying all, so chill everyone. But if you have self or professionally diagnosed ADHD, do you "exercise" and build your focusing muscles regularly by engaging in tasks that require focused attention? Do you discipline yourself away from the junk (doom scrolling, short video consumption)? Or like someone who is a servant of their affliction do you avoid what's hard and indulge in these behaviors that reinforce the problem?
TL;DR: The ability to focus is a perishable skill for everyone.
This doesn't always get a lot of traction, but I have medically diagnosed ADHD. The meds cause terrible and dangerous side effects for me, so I cannot take them.
I decided to actively train my focus in my early twenties. Over time I got better and better at it. I require a few prerequisites: well rested, recently exercised, not too hungry, have no meetings coming up, and keep my phone away. I'm able to engage my hyperfocus at will. Most days it lasts for 4-6 hours straight. I am able to engage it at least once a day.
This took many years to build up. It's fragile - easily prevented by distractions. But with the right set and setting, once it engages it's an incredible tool. I've accomplished things in 4-6 hours that would normally take me weeks.
I wish I could get the benefit with meds, but here we are. Maybe what I've built is the generic crappy equivalent, but it's all I've got and I'm proud I've gotten it this far.
For how I trained it, I started programming on the train ride home. I worked through low stakes programming puzzles from SICP. I found it worked best if I sprinted to the train station and ate something an hour before departure. I would open my laptop and just try to engage the hyperfocus. Many days I couldn't: people talking, train noises, being hungry, etc. Over time though I got better and better at it. Eventually I would be engaging as the laptop screen was opening. To this day ten years later opening a laptop screen engages the hyperfocus if all the rest of the requirements are met. Over a year I finished SICP, doing about one exercise a day. I then picked up another book of homework problems, then another, and another. Over a few years the skill strengthened, I found I could hold it for the drive home from the train station and keep going after.
It's harder at a desk, often I'll burn an hour or two before I can make it click. But once it does I can hold it even when getting up to go get food or going for a walk. I've gone to a cafeteria with coworkers or driven home while still in a "work trance" that picks right back up when I log in again. It's very strange being in hyperfocus on a programming puzzle while talking to someone over food. I can release it, but then I might not be able to engage it again that day so I often leave it running. Rude to my coworkers at lunch, perhaps, but I gotta make money somehow and this is how I do it.
I credit a large amount of my success to this tool. It's fantastic for just getting epic amounts of work done. I can spend hours tracing down a detailed bug, or building out a whole new module. In this trance state reading code feels as natural as reading English. Big transformations make perfect sense in ways I cannot describe outside the effect.
I've used it to write two books, build my own business, and make a living.
I have started using large language models for book recommendations. I can be very specific about what I am looking for and the recommendations are hyper personalized. If you use some sort of tool that pairs LLMs with realtime data like Gemini the results are even better.
LLMs are actually great for book recommendations. Still not quite as good as looking up "HN best book on unit testing" for most topics, but great for niche things.
But, to be honest, my go-to move these days is joining a community and asking a well-regarded expert. I get all my recommendations from readers these days.
It's been reportedly rewritten from scratch like five times, during which time people have not stopped posting claims that it's exactly the same as it was in 2010.
So is this going to be another clusterfuck like 10DLC? I am glad our company stuck with our guts and intentionally decided not to go outbound, but I almost feel bad for the startups that were banking on full outbound.
Interesting juxtaposition in the comments of "I don't answer calls from numbers I didn't recognize because of how scam-prone modern telephony is" with "many startups have seen huge profit opportunity in making outbound phone calls automatically at scale".
Reminds me of the story about overhearing Juul employees on BART talk about how hard they were working to make sure their kids never got anywhere near their product and that if other parents didn't do that, what happened next was their own fault.
Looks like the FCC basically killed outbound AI calling companies like Air.ai, and does not seem to affect inbound companies like ours (https://echo.win)
Interestingly they explicitly mention AI generated voices, does that mean voices generated by traditional TTS engines are fine?
Those voices were already prohibited. This ruling specifically addresses agents "emulating human speech and interacting with consumers as though they were live human callers when generating voice and text messages".
Based on the (alarming) demo on Air.ai's homepage, that sounds like it would be prohibited unless the user consented to be contacted in that manner when providing their phone number.
Our company https://echo.win/ provides inbound phone call automation and management using AI for businesses. Generative voices are going to add a lot of value to our product.
I have had luck with doing this on GPT4 with careful prompting, but GPT 3.5 is pretty reluctant to respond with anything other than straight up conversational answers.
I just saw that Azure OpenAI service has a SLA and OpenAI does not. I thought they would have separated the infrastructure for free ChatGPT users and paying API customers.
When the big earthquake in Nepal happened in 2015, I was working with a volunteer organization called Translators Without Borders to help with translation during relief efforts. Since I was in the USA I could not contribute back physically, so this was the next best thing.
My goal was to help volunteers that were in the field in Nepal communicate in English -> Nepali and back. Even though this was somewhat effective, there was still a communication gap because most people in Nepal in remote parts could not even read in Nepali.
I looked around for solutions but couldn't find any Nepali Text To Speech solutions. The builder brain in me fired up and I decided to build a Nepali Text To Speech engine using some of the groundwork that was laid by Madan Puraskar Pustakalaya (Big Library in Nepal) which they had abandoned halfway.
I spend all night hacking along to build a web app that let the volunteers paste translated text and have it spoken. The result was https://nepalispeech.com/ and the first iteration of this was built in just 13 ish hours.
I hope the people that got affected by the earthquake are in a better situation now.
Sad to see all the non Chinese open source models being at least one generation behind.