ajohannes.bsky.social
Programming: AI, Machine Learning, IoT, automation
Helsinki, Finland
60 posts
36 followers
208 following
Active Commenter
comment in response to
post
I think we should define "fair training". If you can replicate multiple pages verbatim it's not very human and not far from deliberate piracy. If not differential privacy, perhaps reasonable stochasticicity, or one could even design model/training exactly for replication instead of "extrapolation".
comment in response to
post
Mielenkiintoista. Mutta miten näitä kuvia pitäisi tulkita? Esim. vasemmiston keskiarvo ~0.4 kun sukupuolia ei huomioida ja >1.2 molemmille sukupuolille.
comment in response to
post
Being on the cusp of a GPT3 moment sounds like it would still take at least 2 years for Waymo to get where they are now had they chosen the AI way. And even GPTs still aren't good enough for safety critical tasks so I'm not sure. For the late comers, it'll definitely be easier.
comment in response to
post
Too far, but in the right direction. Or is that an apology?
comment in response to
post
With AGI, I guess you could drop the word "researchers" from this dialogue?
comment in response to
post
In Helsinki, while there may be higher fares pub.transport, I think the main reason is simply remote work, visible especially in the age groups 18-44. Car usage hasn't been reduced quite as much, which also makes sense since traffic jams and available parking has been even greater limiter earlier
comment in response to
post
Not quite sure which pieces you refer to, but more than complete denial, I think I've seen arguments that AGI won't appear within a year or even 3 and probably takes more than a decade. Which is also an argument against the current valuations of some AI companies that keep hyping the opposite.
comment in response to
post
No doubt this will make Americans a bit greater, even if living a bit shorter
comment in response to
post
The issue seems to be mentioned in the current version of the article
comment in response to
post
I was about to yell no on your first post, until I read the comments on the other thread. Anyway, works better for some other people, with different kind of followers. It's a bummer you can't choose the ones you have. If only you could post under separate topics... you'd soon have another reddit
comment in response to
post
I think he has a point there. There are far more native Spanish speakers in the Americas so it would make sense for English to be the number 2
comment in response to
post
There is the not so negligible chance that this will be the main CVPR event this year while everything else will be remote
comment in response to
post
I really thought the point was that there are so many fully or 95% LLM written papers where the author (i.e., prompt engineer) never intended to announce the fact.
comment in response to
post
With the test in the pic, I guess it not necessarily just computer vision, but vision in general. Surely nice if the car could have all possible sensors. And in case of plain driving assistance, the more orthogonal set of sensors (vs human) the better.
comment in response to
post
Yeees I guess a descriptor. Having only glanced some of the related papers my mental model of the components is quite vague. And the issue with light/dark points feels so tied to missing the description of scene in the first place...
comment in response to
post
So can you next make a keypoint detector that doesn't just detect points but also describes them in a few embeddings. For others to match the red brick in a gray wall and the gray one some other time?
comment in response to
post
With @jetbrains.com IDEs, it may be sometimes challenging to know what is code and what is some meta info from linters and other tools. And it feels quite ok when you get used to it. I guess one could go even further. I love ruff so I don't need to think about style. Learning a new one takes time
comment in response to
post
Reading this, I think maybe we need to stop formatting python, at all. You could implement editor plugins that display the code any way you want and some other way for others. You could use some format for storage, e.g., a one that is optimum for calculating diffs and git rebase gets way nicer
comment in response to
post
Natsitervehdykset ja moni muu juttu päässyt esille mediassa vasta tammikuun lopulla joten teslan pudotuksen voisi olettaa jatkuvan
comment in response to
post
It would be good to list your terms in the profile bio so that we can understand each other.
For me the climate change is "Massive Atmospheric Gas Anomaly", or MAGA, but I can make a few new ones if they ever get to the other list.
comment in response to
post
Stephen's longer blog post about the subject for those of us who didn't quite get the logic in BSky format: writings.stephenwolfram.com/2018/11/logi...
comment in response to
post
I've thought that it's distillation only if you train against the probabilities, not tokens. Did I get this wrong? OpenAI of course doesn't specify how they train the model but surely they would have the probabilities too for their own model.
comment in response to
post
... it's just a matter of some fine tuning. The company has claimed a cost of 2500-5000$ for a robot. With 1.90€ for delivery and >10 deliveries per day I'd think the company is already making profit on per-robot basis.
comment in response to
post
...but clearly it doesn't always work. They also stop before hitting humans, but only at ~2m distance. And on snowy road they seem to sway left and right like a drunkard, still going 6km/h, faster than the most pedestrians. Their manners seem to frighten some of the elderly nearby, but perhaps..
comment in response to
post
...but looking at the tracks, it had turned back home and switched to cycle lane, probably due to a remote command. They're said to have winter tires but most cyclist wouldn't take such a thin tires on snow. I've seen them stop and reverse when seeing obstacles on the ground, even chunks of snow...
comment in response to
post
They travel up to 4km distances, more than just the last mile, and do so in all weather. Although I think the latter will still cause issues. This one had stopped without clear reason when it had been snowing for 2 hours. The lights went out. Perhaps it was rebooting? After an hour it was gone..
comment in response to
post
So my first BSky post/thread is about the robotization that spreads at surprising speed even in the snowy Helsinki. Starship robots delivering groceries already reach more than half the population (the not darkened areas). Notably, not the most central part of the city.
comment in response to
post
I think it's not just laws that need to change. In a few years someone needs to figure out if dog/humanoid delivery robot needs to buy a ticket to enter tram or subway
comment in response to
post
Here's the same pair with roma. It finds gazillion points in the white building and I had to tune the RANSAC to actually match anything elsewhere. Didn't notice major difference in other pairs before running out of GPU quota. Water seems to match any water. Perhaps that's where the name comes from?
comment in response to
post
I tried their demo at huggingface.co/spaces/Littl... on some of my photos against google earth. Surprisingly difficult to make it match anything correctly. Need to adjust view very close. I guess the GE is a bit too synthetic. Anyway, I'm not super into the scheme, my expectations might be too high.
comment in response to
post
Have they already considered that there's mostly people happy with their current social security? 40+ electors for Canada and a few for Greenland means mostly Democrat presidents for some decades.
comment in response to
post
Look, we're all LLM shills here. If you're not using LLMs to generate content we can blame you for merely creating training data.
comment in response to
post
And here's the message that mentions it sunk. For those who haven't watched bsky for almost an hour due to visit by Santa bsky.app/profile/oale...
comment in response to
post
I yet haven't seen any analysis if o3 is good in Frontiermath because its knowledge or because "intelligence". It's hard to answer without knowing questions but it sounds like it's still largely due to knowledge and the intelligence is poor compared to humans. Especially with the failures on ARC-AGI
comment in response to
post
Taichi is here: github.com/taichi-dev/t... but it's difficult to find who is behind all this, except that maybe it's some Chinese startup. Perhaps there's a language barrier, or a cultural one, but still it would be nice to know the ways they intend to profit
comment in response to
post
After reading about Genesis, I think quite a few features seem to come directly from Taichi ("parallel programming language for high-performance numerical computation") that uses pytorch backend. With all the marketing here, I'm most puzzled I haven't heard about Taichi before
comment in response to
post
I wonder do they dare to cross the English Channel
comment in response to
post
Don't know about OCaml, but I think Rust will have certain benefits when there's large amounts of code that is only poorly, or not at all read by humans, especially compared to non-memory-safe languages
comment in response to
post
Here's one starship ready to face the snow today in Helsinki. They seem to slowly expand delivering groceries. I expect some pause during the worst snowfall
comment in response to
post
Does Isaacman actually own SpaceX shares? I haven't seen any mention of such anywhere. Polaris Dawn program doesn't seem to have any profit making plans either. To me it seems he's intentionally spending money on advancing space tech.
comment in response to
post
How much US defense companies do research on their own, i.e., outside of government programs?
comment in response to
post
Is it known how fast the CIA fact book gets updated on his status if he doesn't die? I guess he could still be the acting Syrian president in Moscow.
comment in response to
post
And what about the case when someone creates a trillion message open dataset from your server?
comment in response to
post
I mean, agents often do feel more marketing that technical terms. So a proper suit is a nice add-on.
comment in response to
post
Wears black tie and sun glasses
comment in response to
post
I'm thinking should you rather use something like gifts but I'm not sure would it cause huge number of tiny commits for almost every line github.com/presslabs/gi...
comment in response to
post
Not sure I see the difference. Only the user knows who is using the key. Reading the snum requires physical access and special tools. IP and email are probably easier to fetch.
Web sites have learned to cope but I'm afraid if you dig deep enough my fav cafe has to ask to sign a ToS before entering
comment in response to
post
..which should be quite low entropy data since the locks didn't have clock-chips and nobody except the user knows for sure who is using the key. Only the use of some more rare (e.g. master key) should increase the entopy.
I think if you look closely fineweb etc datasets will break all the same rules