dlevine

A few years back, I replaced all the bedroom light fixtures in my condo with fan/light combos. It has been getting hotter in SF, and I thought fans might help during the limited number of hot summer days. I didn’t have room for both a fan and a ceiling fixture, so I went with combo units.

I actually ended up with one fan that looks exactly like the enclosed one (I had limited clearance and needed something small). It works fine as both a light and a fan. If the bulb goes out, I can just replace the whole thing (it wasn’t that expensive).

For the other one, I bought something higher end bc I had enough clearance to go with larger blades (44” Minka Aire Light Wave). The light is a plug-in module, so if it goes out, I can find a replacement (assuming they still make the module).

Overall, these combined units aren’t perfect, but they let you add a fan without needing to add more wiring. The one thing to be careful of is to make sure your electrical boxes are fan rated.

Apparently 2 stones is a huge advantage. An estimate is that the computer is roughly 4-600 ELO stronger on an even match.

Also, the human played a strategy tailored to that huge initial advantage. He said that the AI did not handle this particularly well, and played high probability moves instead of trying to lure him into a mistake.

Also, even though this was the best Go engine, it was not running on a supercomputer, and had a relatively limited amount of time per move.

So, this was an important victory for a human, but not a sign that humans are now stronger than AIs at Go.

It’s not an either/or. It’s doing both. The video makes it easier to understand what the code does. We have used Looms on PRs since before AI, and agents are beginning to be able to do this on PRs they generate.

I think this is more a symptom of the problem than the actual problem.

The issue is that we can generate tons of code using AI, but then are blocked on having humans review all of it.

I don’t think we should auto-approve all of this code without human review - that clearly doesn’t work either.

What I do think we need is probably at least two-fold 1) better ways to explain these big PRs to human reviewers. 2) better ways to verify the functionality of a piece of code. Things like auto generating walkthrough videos

I’m not sure that even this is enough. I’m sure there will be agents that try to solve this problem.

I had a TI-81 (and then an 85 later on). I wrote a blackjack program. There was no link cable for the 81, so other kids had to type it in manually.

Subpixel rendering only gives you 3 effective pixels. M and N aren't really readable with less than 5.

It seems to have worked for Bill Gates as well. He definitely did some not so nice things when starting and running MS - I think it unfortunately goes with the territory of running a successful company at scale. But subsequently he has become more know for his philanthropy.

When I was in the college in the early 2000s, I had a friend who ran OpenBSD. He always sang its praises, mostly because it was the most secure operating system.

I tried a bunch of Linux Distributions and FreeBSD before mostly settling on MacOS, but never actually got around to running it.

Glad to see OpenBSD is still being actively developed.

Something I find weird is that this article compares a 9950x with two different laptop CPUs and concludes that performance has increased massively in the past few years. If you compare the 9950x with its two Desktop predecessors (released 2 and 4 years before), you see about a 6% increase from the 7950x and a 45% increase from the 5950x. So you should consider upgrading regularly, but potentially not every single generation. I think it makes sense to consider the performance and offer an upgrade when you see a 50% or so cumulative improvement. Everywhere I have worked has upgraded developers every 3-4 years, and it might make sense to upgrade if there is a massive change (like when Macbooks went to M-series).

As for Desktop vs Laptop, that is relevant too. Desktops are typically much faster than Laptops because they are allowed much larger power envelopes, which leads to more cores and higher clock speeds for sustained periods of time. However, there is always a question as to whether your use case will be able to use all 16/32 cores/threads in a 9950X CPU. If not, you may not notice much difference with a smaller processor.

Source for CPU benchmarks: https://www.cpubenchmark.net/compare/6211vs5031vs3862vs5717/...

I haven't done this that much, but have found it to be pretty useful.

When it just gives me the answer, I usually understand but then find that my long-term retention is relatively poor.

If someone uses AI to generate an output, that should be stated clearly.

That is not an excuse for it being poorly done or unvetted (which I think is the crux of the point), but it’s important to state any sources used.

If i don’t want to receive AI generated content, i can use the attribution to filter it out.

My theory is that, when done properly, it’s much closer to science than engineering.

And by “done properly,” i mean done in a regimented way with evals to verify that a wide range of inputs produce the desired outputs.

Prompting is much closer to discovering the properties of an already existing system than building something using engineering methods.

I had a licensed copy of Executor back in the mid-90s. It was the coolest thing ever. Thanks for being one of my inspirations to go into software development.

I think of LLMs as knowing a lot of things but as being relatively shallow in their knowledge.

I find them to be super useful for things that I don't already know how to do, e.g. a framework or library that I'm not familiar with. It can then give me approximate code that I will probably need to modify a fair bit, but that I can use as the basis for my work. Having an LLM code a preliminary solution is often more efficient than jumping to reading the docs immediately. I do usually need to read the docs, but by the time I look at them, I already know what I need to look up and have a feasible approach in my head.

If I know exactly how I would build something, an LLM isn't as useful, although I will admit that sometimes an LLM will come up with a clever algorithm that I wouldn't have thought up on my own.

I think that, for everyone who has been an engineer for some time, we already have a way that we write code, and LLMs are a departure. I find that I need to force myself to try them for a variety of different tasks. Over time, I understand them better and become better at integrating them into my workflows.

I’m job searching right now, and it’s definitely not the doom and gloom that I’m hearing about.

I think it’s a challenging environment for developers who are either inexperienced or who have skills that are out of date. I have found that companies are a bit more picky than i remember about knowing the exact tech stack they use, but they are still making offers and those offers are pretty good.

Note that I’m applying mostly to mid-sized non-public companies. I’m not sure what it’s like applying to MAANG-types right now.

I think of LLMs like smart but unreliable humans. You don't want to use them for anything that you need to have right. I would never have one write anything that I don't subsequently go over with a fine-toothed comb.

With that said, I find that they are very helpful for a lot of tasks, and improve my productivity in many ways. The types of things that I do are coding and a small amount of writing that is often opinion-based. I will admit that I am somewhat of a hacker, and more broad than deep. I find that LLMs tend to be good at extending my depth a little bit.

From what I can tell, Sabine Hossenfelder is an expert in physics, and I would guess that she already is pretty deep in the areas that she works in. LLMs are probably somewhat less useful at this type of deep, fact-based work, particularly because of the issue where LLMs don't have access to paywalled journal articles. They are also less likely to find something that she doesn't know (unlike with my use cases, where they are very likely to find things that I don't know).

What I have been hearing recently is that it will take a long time for LLMs will be better than humans at everything. However, they are already better than many many humans at a lot of things.

The name SheepShaver is a play on ShapeShifter, which was a Mac II emulator for Amiga. I remember running an Amiga Emulator (UAE) AND Shapeshifter on top of that since it was the best Macintosh Emulation at one point in the late 90s.

I have been playing around with MCP, and one of its current shortcomings is that it didn’t support OAuth. This means that credentials need to be hardcoded somewhere. Right now, it appears that a lot of MCP servers are run locally, but there is no reason they couldn’t be run as a service in the future.

There is a draft specification for OAuth in MCP, and hopefully this is supported soon.

Non-Paywalled Link: https://medium.com/@dana11235/feeling-the-vibe-coding-with-a...

I have found feature flagged rollouts to be one of the biggest advances in fairly recent software development. Probably too much to say about it in a comment, but they massively de-risk launches in a number of important ways, both in being able to quickly turn a feature off if it has unintended consequences and being able to turn it on for a very specific set of users.

With that said, I think that LaunchDarkly and the like are a bit expensive and heavyweight for many orgs, and leaving too many feature flags lying around can become serious debt. It totally makes sense to start with something lighter weight, e.g. an env var or a quick homegrown feature in ActiveAdmin.

About a year later, I got the P3-550 that overclocked to 733. Not quite as good of an overclock in terms of percentages, but I ran that machine for 5 years with no issues.

I played Disco Elysium when it came out and enjoyed it. In particular, I thought the Inland Empire skill was pretty awesome. I can't imagine what the game would be like without the ability to talk to inanimate objects.

This is super impressive! It's a cool POC, although it is already clear that it would be feasible for Apple to put M.2 slots in Macbooks if they wanted to.

I wonder how much it would cost to have someone replace the BGA NAND chips in my Macbook. Apple charges $6-800 for a 2TB upgrade for a Macbook (depending on whether it's a 250 or 500GB drive originally). Someone would have to be able to do it for like $2-300 for it to be a feasible upgrade, especially considering that my warranty would be void. I assume there are people overseas who could do it cheaply. I assume it would be fairly quick for someone who knows what they are doing.

AI adoption can be increasing even if progress of the base technology has slowed down. LLMs as a technology are very new, and there are doubtless tons of interesting uses we have yet to discover.

California’s state tree is the redwood, and that’s where their HQ is.

Reading this, I do get the impression that the Partially Meets Expectation was given to get more work out of this guy. It's the worst kind of manipulation - spinning something that was an unqualified success into a "failure." Especially because the boss clearly already had an internal narrative, and it sounds like they cherry-picked feedback from coworkers to support that narrative.

Having had something somewhat similar happen in the past, it seems like this is common at high-growth tech companies. Everyone pretends to be your friend, and there are definitely some genuine people, but getting ahead is clearly number 1.

I'm surprised they are still shipping these with 256GB of storage base. I had a Macbook with a ~500GB SSD in 2012 (I installed it), and a 500GB spinning disk in like 2008 (also user installed).

A 500GB SSD can be had for <$50 these days, and a 1TB for <$100. Still plenty of profit for Apple, even if they bump up the base storage to 512GB and make 1TB a $200 upgrade...

The specialization did a good job of explaining all of the basic concepts of RL in a fairly understandable way. The instructors don't assume a huge amount of prior knowledge, and make it pretty accessible. They split reasonably between theory and applications. I would say that you need some background in math and ML.

The courses do build upon each other. You could do just the first course and get a good overview of RL without diving into the details, but I don't think you could start midway through the specialization.

I would also suggest looking at the Hugging Face Deep Reinforcement Learning course. That one is very different (focused more on application than diving deep into the theory), but it's taught by a non-academic and really tries to explain the concepts in a way that is approachable to most programmers.