I’ve used AI by just pasting code, then asking if there’s anything wrong with it. It would find things wrong with it, but would also say some things were wrong when it was actually fine.

I’ve used it in an agentic-AI (Cursor), and it’s not good at debugging any slightly-complex code. It would often get “stuck” on errors that were obvious to me, but making wrong, sometimes nonsensical changes.

There’s a lot of indication that LLMs are peaking. It’s taking exponentially more compute and data to get incremental improvements. A lot of people are saying OpenAI’s new model is a regression (I don’t know, I haven’t really played with the new model much). More foundational breakthroughs need to be made, and these kinds of breakthroughs are often the result of “eureka” moments which can’t be manifested by just throwing more money at the problem. It’s possible it will take decades before someone discovers a major breakthrough (or it could be tomorrow).

China leads the world in scientific publication, even when only taking into account reputable journals and high-impact publications. There’s no doubt in my mind the US will decline further with the current attacks on science and education, and anti-intellectualism in general.

In the grand scheme of things, I think AI code generators make people less efficient. Some studies have come out that indicate this. I’ve tried to use various AI tools, as I do like fields of AI/ML in general, but they would end up hampering my work in various ways.

I almost never have cash or even change now, so that’s what I tell them. I used to give them some money if I had it and wasn’t immediately going to use it. One of my old friends used to give them a beer out of a pack he bought if there was a person outside the store or on the corner begging on the drive to wherever he was going.

I had a Mox and it was the most stable router I’ve ever had. Never had to power cycle it, and I even have to power cycle my fancy Asus somewhat often.

Newer vehicles typically have ceramic stock (well at least my newest car did), but I think “organic” were standard until relatively recently. There’s pros and cons to each.

Yeah, Apple’s stock went up 5% yesterday because of Tim Cook’s meeting with Trump (where he promised $600 billion investment in US manufacturing), and Apple’s saying they won’t be affected by the new 50% tariffs on India. There are also ghost factories that were built during Trump’s first term.

I never understood the popularity of AWS. It’s much cheaper using VPSs and even dedicated servers sometimes. I’ve worked on very cost-sensitive projects where I rolled our own highly-available k8s and postgres clusters on dedicated servers and VPSs and saved the company a shit load of money. Only used the “cloud” to store backups (Backblaze). There’s tons of other options other than major “cloud” providers, and they’re often much cheaper.

Hmm, a lot of my career was done doing embedded programming, where mistakes in production are very costly, and software/hardware has to be released with basically zero bugs, so that may be where the disconnect is. I still think bugs and technical debt are costly elsewhere too if the product is going to have a long lifecycle, but executives are just dumb.

For open source stuff, Codeberg is good. For private stuff, just git + ssh is good. Gitlab and Bitbucket are fine for corporate stuff, I guess. An organization could just self-host a Forgejo (or Gitlab) instance as well.

I keep hearing stuff like this, but I haven’t found a good use or workflow for AI (other than occasional chatbot sessions). Regular autocomplete is more accurate (no hallucinations) and faster than AI suggestions (especially accounting for needing to constantly review the suggestions for correctness). I guess stuff like Cursor is OK at making one-off tools on very small code-bases, but hits a brick-wall when the code base gets too big. Then you’re left with a bunch of unmaintainable code you’re not very familiar with and you would to spend a lot of time trying to fix yourself. Dunno if I’m doing something wrong or what.

I guess what I’m saying is that using AI can speed you up to a point while the project accumulates massive amounts of technical debt, and when you take into account all the refactoring and debugging time, it results in taking longer to produce a buggier project. At least, in my experience.

I think part of it is because they think they can train models off developers, then replace them with models. The other is that the company is heavily invested in coding LLMs and the tooling for them, so they are trying to hype them up.

Chamoy on fruit or candy is awesome.

I’ve tried Copilot for a while and played around with Cursor for a bit. I was better and faster without Copilot due to sometimes not paying enough attention of the lines it would generate. This would cause subtle bugs that took a long time to debug. Cursor just produced unmaintainable code-bases that I had no knowledge of, and to make major changes, would be faster for me to just rewrite it from scratch. The act of typing gives me time to think more about what I’m doing or am going to do, while Copilot generations are distracting and break my thought processes. I work best with good LSP tooling and sometimes AI chatbots (mostly just for customized example snippets for libraries or frameworks I’m unfamiliar with; though that has its own problems because the LLMs knowledge is out of date a lot) that don’t directly modify my code.

People have different levels of “nerves” as others, and it kind of sounds like you may filtering out applicants on an arbitrary metric (how nervous a person may be in an interview). Don’t have enough information about your process to say for sure (obviously), but it may be something to think about. Interviews can be very high-stakes for some people (such as “I may become homeless”), and not for others (“my parents are rich”). After hired, it’s not necessarily as high-staked, and toy problems aren’t what SEs work on day-to-day.

If doing an overlay network (network on top of the Internet), you probably won’t be able to do much better than Tor or i2p.

We confirm the trilemma that an AC [anonymous communication] protocol can only achieve two out of the following three properties: strong anonymity (i.e., anonymity up to a negligible chance), low bandwidth overhead, and low latency overhead.

https://freedom.cs.purdue.edu/projects/trilemma.html

This applies to all types of anonymous networks as well (BT, Wifi, etc).

The Dems will not be able to repeal the tariffs if they only control the House. It currently looks very unlikely they’ll control the Senate, and it’s just a coin toss right now if they’ll even control the house (will probably turn unlikely after Texas redistricts).

More like $200k in total comp (but I suppose if the stock rockets like 5x after you’re hired, it can end up being $400k). Senior positions can make > $400k/yr in total comp. Some companies have back-loaded vesting schedules so they can get rid of you before the majority of your options vest though.

As I understand it, pedophilia is just attraction; not taking action. And many people who were abused as children themselves end up developing the condition. I think it is treatable, but probably not “curable” (maybe, IDK).