Horizontal gene transfer is looking a lot more common than we learned in biology class. Fungi have a class of proteins (Starships) which transfer hundreds of genes, beetles picked up cellulose-digesting proteins from bacteria in their digestive tract and embryos.
> One of my AirTags is inexplicably 1600 miles away in Texas. It's one of my floating ones that I throw into bags when traveling as needed.
> Maybe it dropped out of my bag only to be found by someone who thought it'd be fun to confuse the owner. Or maybe it became e-waste to be dumped in another country.
Well that's sort of the point, does it have a positive effect for that cohort of kids years later? The OLPC hardware was looking rough several years ago and Uruguay's Plan Ceibal and other projects aren't using their hardware today.
For reasons unclear to me, Sugar Labs (their OS / UI) just got nonprofit status last year?
Well we have the idea and the next best thing to official code, but if this was a big revelation where are all of the Titan models? If this were public, I think we'd have a few attempts at variants (all of the Mamba SSMs, etc.) and get a better sense if this is valuable or not.
Further, if I use AI-written citations to back some claim or fact, what are the actual claims or facts based on? These started happening in law because someone writes the text and then wishes there was a source that was relevant and actually supportive of their claim. But if someone puts in the labor to check your real/extant sources, there's nothing backing it (e.g. MAHA report).
Well it's cool that they released a paper, but at this point it's been 11 months and you can't download a Titans-architecture model code or weights anywhere. That would put a lot of companies up ahead of them (Meta's Llama, Qwen, DeepSeek).
Closest you can get is an unofficial implementation of the paper https://github.com/lucidrains/titans-pytorch
Humans have two copies of each chromosome. Plants in nature can have many more, and some can be induced to duplicate their chromosomes which sometimes gives them larger seeds and resilience to environmental conditions. Similar to hybrid crops but somehow containing the full genomes of both parents. Bread wheat is hexapoloid (containing sub-genomes from three varieties of wheat) and quinoa is tetraploid (containing two different species). There have been projects for ~100 years to make polyploid rice with heavier grain-weight, but they haven't been able to reproduce.
IBM challenged that the 2019 case could be handled by a supercomputer [1].
The main issue is that these algorithms where today's early quantum computers have an advantage were specifically designed to be demonstration problems. All of the tasks that people previously wanted a quantum computer to do are still impractical with today's hardware.
I am on the other side of this problem and was surprised because it is very easy to contact me. When the other person with my name forwarded the emails, it was all careless and unwanted recruiter mail. Someone goes into work, types first.last@gmail, and hits send without doing one Google search. Incredible.
An annual 0.1% chance of one casualty (so: a passenger jet less than 1 in 1,000 years) doesn't become significantly more likely in two years. Also I assume the FAA actuaries would have forward projections of Starlink launches.
Maybe this is space shuttle math where real-world accidents tell us that the risk is significantly higher. But it'd be the first documented case of a meteor or space debris, so I'd guess it's still unlikely.
Like LLMs, this isn't the sort of thing where a small group would make a sudden advancement and be made secret, and I doubt that the NSA can make theirs significantly faster than any industry team today. I think more likely you would need to get worried if someone got one scalable to hundreds or thousands of logical qubits and then stopped publishing.
I agree this makes little sense for Cloudflare to jump on the crypto bandwagon now. Maybe they want to retain some talent by turning this into an official project.
Is the premise that it makes more sense for an AI agent to pay in prepurchased stablecoin tokens instead of direct access to a credit card?
I really appreciated her speaking to young people, even riding the NYC subway for the first time to record "Subway Takes" last year: https://www.youtube.com/watch?v=CAkwo6JPV00
Google Glass has been a thing for 12 years, Snapchat's glasses have been around for 9. And you rarely see someone using them. (I'm aware of reports of ICE officers wearing the Meta glasses, but this is probably an extension of norms around police bodycams)
As I understand it, Siri and Alexa could be plugged into an LLM, but changing it to an "open world" device that can tell your kid something disturbing, text all of your contacts, buy groceries, etc. comes with serious risk of reputational harm. While still falling short of people's expectations if it isn't ChatGPT-quality. OpenAI is new enough that they get to play by different rules.
Everyone keeps asking this but AFAIK they're an alternative database host. If you're a company you can compare their pricing and availability to RDS and other companies. If you have an open AWS contract, or a hobbyist developer who already has databases on AWS you might not see any reason to switch away.
I was putting together an essay/podcast about a news event in another country, and I was able to collect multiple YouTube videos, search and translate their transcripts, and ask questions like 'is this the video where someone talks about _?'
I think it's from people who are programmed from early e-commerce days to think using their credit card online is an extreme risk, and that Paypal is shielding them or insuring their purchase in some way.
That said, I know some small nonprofits where that's their preferred way to donate online.
I work at a university and we do a lot of moving data around, inspecting the files and columns, scripting, etc. so we have everyone use QGIS. Governments and other major consumers have open-ended long-term contracts for whatever Esri products they can think of, so those are solid.
For me the real ongoing question is the role of MapBox, MapLibre, to some extent Google Maps API, and other web-first solutions. It's difficult for Esri to connect with the average web developer or researcher who just wants to start with clickable pins on a map.
'npm install' and 'pip install' can both run scripts on your computer. Both ecosystems have this risk and loose monitoring, so there are days where packages are messed up. I don't think you can avoid malware by picking one over the other.
There was a proof-of-concept paper about buying up expired domains in the LAION image dataset and poisoning multimodal LLMs that way (then LAION was just a list of image URLs). As I understand it, the paper was exaggerating its reach, and LAION has newer versions, torrents, etc.
I don't think enough people talk about how lucky we are that the International Date Line can roughly follow a line of longitude 180 degrees from the Greenwich meridian. If you were on a planet with more land, or we could only draw it through the Atlantic or Australia, timezones would be a lot weirder.
People are not reading and absorbing the longer pieces of writing. I worked at a company which would create information-dense, wordy PowerPoint slides for client updates, and at some point you see it has no point besides trying to impress. If we actually needed something in an email, they told us to put it in the first sentence.
I've also noticed it when people post LLM writing on Reddit. Something may give me pause, and then re-reading the content any given paragraph was way off. I had even glossed over the bolded conclusion "it’s a coaching-wheels moment" (?) because as you read it your brain thinks of a way it could make sense.
It is a command line tool "huggingface-cli", it's just installed with a differently-named pypi package. I wouldn't call this a full hallucination because anyone could make this mistake.