hmottestad

Good for you. Cross my fingers that you'll land a good job soon. Or create your own job.

Might be related to EFTA.

My mum is using my old 12 Pro, which is about the same age as yours. She’s still happy with it, though to me it feels old.

I’m sure the CPU designers would love it if they didn’t need several different layers of cache. Or no cache at all. Imagine if memory IOPS were as fast as L1 cache, no need for all that dedicated SRAM on the chip or worry about side channel attacks.

Does SQS guarantee delivery to all clients? If it does then that’s doing a lot of heavy lifting for you.

If it doesn’t guarantee delivery, then I believe you will at some point have a client that reads a cached value thinking it’s still valid because the invalidation message got lost in the network.

I was looking for one to use for named entity extraction and found this fine tune here: https://huggingface.co/dslim/bert-base-NER?utm_source=chatgp...

Its only 108 million params.

I definitely have a homelab, though I’ve tried to keep mine concealed and out of the way. It’s fairly reliable, but so overkill I wouldn’t want to talk to loudly about it in fear that people would burst my tiny little bubble of justification.

I do have a really cool washing machine from Miele that has a cartridge system in the bottom so that I don’t have to measure out detergent by hand. Just press a button.

But I’m not OP.

I wouldn't equate model weights with source code. You can run software on your own machine without source code, but you can't run an LLM on your own machine without model weights.

Though, you could still sell the model weights for local use. Not sure if we are there yet that I myself could buy model weights, but of course if you are a very big company or a very big country then I guess most AI companies would consider selling you their model weights so you can run them on your own machine.

Would be very interested to read more about that. Do you have a link?

Mistral isn’t using misleading benchmarks. I linked to DeepSeek’s own benchmark results that DeepSeek created. I couldn’t find anything newer.

Can you link me to the benchmark you found?

With how amazing the first R1 model was and how little compute they needed to create it, I'm really wondering how the new R1 model isn't beating o3 and 2.5 Pro on every single benchmark.

Magistral Small is only 24B and scores 70.7% on AIME2024 while the 32B distill of R1 scores 72.6%. And with majority voting @64 the Magistral Small manages 83.3%, which is better than the full R1. Since I can run a 24B model on a regular gaming GPU it's a lot more accessible than the full blown R1.

https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-...

760 600 queries per day. That’s about 8.8 queries per second.

Per user it’s about 15 queries per day. I’m sure there will be some that are incredibly active and some that aren’t active at all, but 15 per day seems quite reasonable.

One thing that would be great is if Codex would check the CI status for the PR it created and see if something didn’t pass. That way it could actually be a bit more like a human.

I’ve been testing on a large open source Java project that I contribute to.

Took a lot of trial and error to get the setup script working. If commands were too verbose then everything would crash, and if it’s not outputting anything for a while then the script is cancelled because it thinks that it’s stalled.

There was also the issue with maven not detecting the default Linux http proxy.

I’ve also put a lot of effort into trying to create a good AGENTS.md file.

So far it’s been very slow going, but I have managed to get it to start porting some of my unit tests to use test containers for running up Elasticsearch.

This particular JEP is just: "Change compact object headers from an experimental feature to a product feature."

Very good performance results though. Particularly like the json parsing benchmark showing a 10% performance improvement.

You can specify a startup script for your environment in the Edit -> adbvaned section. The code placed there will run before they cut off the internet access. Also worth noting that it uses a proxy stored in $http_proxy.

Took me an few hours today to figure out how to install maven and have it download all the dependencies. Spent an hour trying to figure out why sudo apt-get update was failing, it was because I was was using sudo!

Crawling the web is costly. I assume it's cheaper to use the results from someone else's crawling. I don't know what Kagi is using to argue that they should have access to Google's indexes, but I'd guess it's some form of anti trust.

Dunno about Redis, but for Elastic I still feel sorry for them being thrown around like a rag doll by Amazon. On principal I will not use the Amazon fork, because I don’t want to support a company that would prefer to fork a project rather than fork over some cash. Amazon is more than willing to sell you their Elasticsearch fork at a loss as long as they can eventually recoup the losses when Elastic inevitably dies. At which point they will naturally abandon the open sources side of their fork and continue development in private, at a much slower rate, while doubling the price of the AWS service. At which point you’ll have no choice but the pay up, cause there aren’t any competitors left.

Tried it with o1-pro:

> You’ll find that the actual face of the coin under the cup is tails. Seeing “heads” in the mirror from underneath indicates that, on top, the coin is really tails‑up.

What would Microsoft do if I forked their repo, removed all the licenses and then held talks at conferences about my amazing new tool?

Pretty sure their legal department would have my fork obliterated from the face of the earth and I would be crossing my fingers that all I got was a cease and desist letter instead of a lawsuit in Texas.

In the OpenAI demo of codex they said that it’s sandboxed.

It only has access to files within the directory it’s run from, even if it calls tools that could theoretically access files anywhere on your system. Also had networking blocked, also in a sandboxes fashion so that things like curl don’t work either.

I wasn’t particularly impressed with my short test of Codex yesterday. Just the fact that it managed to make any decent changes at all was good, but when it messed up the code it took a long time and a lot of tokens to figure out.

I think we need fine tuned models that are good at different tasks. A specific fine tune for fixing syntax errors in Java would be a good start.

In general it also needs to be more proactive in writing and running tests.

Don’t know about neovim, but the profiler and debugger in IntelliJ are amazing. Couldn’t live without them.

Tomorrow they are releasing the open source GPT-1.4 model :P

Are those with «thinking» or without?

No-name Android TV box is definitively spying on you: https://www.malwarebytes.com/blog/news/2023/01/preinstalled-...

If you go the route of an Android box then be careful with what brand you buy. A couple of years ago a lot of no-name ones came with malware installed.

https://www.malwarebytes.com/blog/news/2023/01/preinstalled-...

We all know how easy it would be to add an option to hide a profile, set one as default or add a «Yes, I do actually agree to watch R-rated stuff and stop asking me about it!».

Feature creep and scope creep are real, but if you want to add a feature for a kids profile then put «disable kids profile» as part of the scope for that task.

As someone with kids I find it really useful to have kids profiles. Netflix one is particularly good, I can configure a general age group and block individual shows.

Our national TV app comes in two flavours, a general app and a specific one for kids. Makes it easier to deploy features for kids without adding any nagging popups for adults.

I think it depends a lot. I remember working in retail for a summer and saw some of the prices. If you wanted to buy an alarm clock, that was 100% markup, but if you wanted to buy the Garmin GPS then it was 15% markup.

I would think that specialised and expensive shoes have less markup than cheaper and more common shoes. But if the cheaper and more common shoes become 50% more expensive then there aren’t really any cheap shoes left to feed the bottom line…

If you wanted to commercialise this, then you could do it on a feature basis.

Create a paid version where you add more features, like support for views.

Found this in an article linked to by this one:

   Exposure to air pollutants increases our risk of developing a range of diseases. These diseases fall into three major categories: cardiovascular diseases, respiratory diseases, and cancers.

    It makes sense to think of these estimates as ‘avoidable deaths’ – they are the number of deaths that would be avoided if air pollution was reduced to levels that would not increase the risk of developing these lethal diseases.