How do those companies make money? Qwen, GLM, Kimi, etc all released for free. I have no experience in the field, but from reading HN alone my impression was training is exceptionally costly and inference can be barely made profitable. How/why do they fund ongoing development of those models? I'd understand if they release some of their less capable models for street cred, but they release all their work for free.
Chinese companies don't always operate on purely capitalistic principles, there is sometimes government direction in the background.
For China, the country, it's a good thing if American AI companies have to scramble to compete with Chinese open models. It might not be massively profitable for the companies producing said models, but that's only a part of the equation
China seems to combine the best points of capitalism (many companies taking many shots on goal, instead of the eastern bloc way of one centrally-mandated solution that either works or not) with the best points of communism (state-sponsored industries that don't have to generate a profit, for the glory and benefit of the state).
Ostensibly, a mix of VC funding and that they host an endpoint that lets them run the big (200+GB) models on their infrastructure rather than having to build machines with hundreds of gigs of llm-dedicated memory.
But on inference they have to compete with other inference provider that just has a homepage, a bunch of GPUs running vllm and none of the training cost. Their only real advantage are the performance optimizations that they might have implemented in their inference clusters and not made public
Adjacent to it are PR reviews. Suggesting simpler approach in PR almost always causes friction: work is done and tested, why redo? It also doesn't make a good promotion material: keeping landscape clear of overengineered solutions is not something management recognises as a positive contribution.
Depends on the management and whether they're involved in coding. Any engineering manager, architect, senior / lead developer etc should appreciate lower complexity.
Of course, if it's the person in charge introducing said overengineering there is a problem.
they can recognise on the informal level, but you can't put it into end of the year review document. What it will be? "Kept N PRs from introducing cruft into our systems?". Fixing or building things is much more visible, than just maintaining high standards.
Worse, to suggest a simpler approach checking existing products/APIs or even preparing toy prototype is required to be confident in own advice. This hidden work is left entirely unnoticed even by well meaning managers/engineers: they simply don't know if you knew or had to discover simpler solution.
You could make same argument in "information superhighway" days, but it turned out to be the opposite: no company monopolised internet services, despite trying hard.
With so many companies in AI race it is already pretty competitive landscape and it doesnt seem likely to me that any of them can build deep enough moat to come ahead.
a few? all sorts of websites and services are thriving on the Internet even after significant consolidation of attention social media caused. Not even close to a dystopian picture parent comment paints.
For example add temporarily nullable column to a large table, deploy new code which starts writing to the new column, in background populate that column for existing rows in batches and finally alter column to be mandatory non-nullable.
Another example of non-trivial schema management case is to make schema change after new version rollout completes: simple migration at the start of the container can't do that.
It must be a solved problem, but I didn't see a good tool for it which would allow expressing these imperative changes in a declarative way which can be comitted and reviewed and tested along the app code. It is always bunch of adhoc ugly scripts on a side and some hand waving deployment instructions.
I tend to prefer to hand-roll schema migrations... but I use grate[1] for the most part. That said, I've created similar tooling for different scenarios.
> They don’t understand it and think it will replace them so they are afraid.
I don't have evidence, but I am certain that AI replaced most of all logo and simple landing pages designers already. AI in Figma is surprisingly good.
I doubt it, you’ll still need humans to create novel ideas and designs because things will get stale after a while and trends/styles will continue to evolve.
Exactly. People are getting very good at detecting AI-generated designs -- because everyone can play around with it themselves and see in what ways they always tend to look alike.
To make an impression, it will become even more important to go with a real designer who can work in creative ways to regain people's attention.
But I have little doubt that a lot of the bread-and-butter, not-too-important, I-just-need-to-have-something jobs will no longer be contracted to actual designers.
Just bought a digital piano as New Year present to myself. So far I am doing single-note-a-time melodies from my child's practice book and so far enjoying my slow progress, but I am really struggling. Spent 3 evenings on a single simplest song and still can't play it end to end reliably.
I have to sign every note with a letter in a music book because only other way to "read" music sheet for me is count lines for each note, which is unbearably slow.
I wonder if there is any modern (AI, bluetooth midi app, etc) way to get over initial hurdles easier?
That's good question. While I think my app might help with your ability to pick out notes - it's not going to assist with stuff like proper pianistic ergonomics, fingering, etc.
I’ve heard some people say that the Piano Adventures Player app is a nice little tool because it serves as a supplement to the books themselves.
Funnily, I felt the opposite when I learnt to read music a few years back - that even though I was extremely slow, I could now (theoretically) read and learn to play any music, just slowly! You'll get faster and better with practice, as with everything. I'm still slow. Just not as slow. But crucially I'm fast enough that it's not the bottleneck anymore, getting the notes under my fingers is.
Former band kid who also just got a digital keyboard. Ime learning to read the staff just came from putting in the time on the instrument, but I’m also looking for ways to speed that up. I had the idea of making flashcards and even putting it into an SRS like Anki to see if I can make the process of (re-)learning the staff faster and make it stick. If you come across anything that would help I’m interested too!
I've found that a colorful note guide (such as: https://www.amazon.com/dp/B0BC8NVW4Q) to be very helpful. Without it, I felt completely lost when looking at sheet music.
> As long as the vehicle has a known starting point, quantum accelerometers can continue tracking its movement accurately, regardless of what’s happening above the atmosphere.
So this is a much more precise inertial guidance system replacement? If true, I'd expect UK MoD to be involved to the point of making technology a military secret, but clearly it didn't happen.
given that there is no dev mode or ssh server running on a console, how do they even read low level binary code such as boot loader? Do they transplant memory chips?
In this case, by using fault injection to induce a glitch into a test mode which bypasses secure boot and loads code from SPI, combined with a SPI emulator (and I2C to send the boot vectors).
Chip-off is a common way to retrieve the ROM of embedded devices. It often requires multiple chip-off reads and a reconstruction of the striped data across the chips.
reply