The article says "We plan to release substantially larger improvements to our models in the coming weeks."
Sonnet 4 has definitely been the best model for our product's use case, but I'd be interested in trying Haiku 4 (or 4.1?) just due to the cost savings.
I'm surprised Anthropic hasn't mentioned anything about Haiku 4 yet since they released the other models.
After my Pebble I tried an e-ink "Watchy" from SQFMI (https://watchy.sqfmi.com/) thinking that the battery life would be great but the battery only lasted a few days.
I've been wearing a Bangle 2 (https://www.espruino.com/Bangle.js2) which feels closest to recreating my Pebble. It has super long battery life and feels a lot like my Pebble did, but doesn't have the polish of the pebble UI and animations.
I've never worn a pebble, but I also have a banglejs 2 and I really love the watch for it's hackability.
I've written my own watchface and a couple of other apps and made changes to a number of existing apps, it's really simple because you can always test your code changes live on the watch while keeping it on your wrist. There's an IDE that connects to it using Bluetooth and the code can be modified during runtime
Lthere's also a great community of hackers and tinkerers that steadily improve the watch.
It might not have the same polish as the pebble had, but it makes it up in hackability. I can only recommend getting a banglejs2 (battery life is also pretty great, I get about 10 days with regular use)
Huh. I appreciate what they did with the tech there, but looking at this Bangle.js the first thing that comes to mind is I hope NuPebble(?) doesn't adopt that excessively-curvy-rectangle shape that screams Apple Watch, I've learned to recoil in disgust even seeing that shape.
"Worldwide, most radiosonde observations are taken daily at 00Z and 12Z (6 a.m. and 6 p.m. EST)" - NOAA.gov
I know we get more weather data from other sources, but it seems insane that these 2 launch times per day (per balloon location) are what make up most of our current weather forecasting data.
You mentioned solar. Do you have the capability (or plans) to run these over night as well?
You’re right - Satellite data like GPS occultations and aircraft derived data still make up most of the data by volume. Soundings remain the ideal data source and are weighted very heavily in models, punching way above their volume.
As of right now power constraints mean we maintain tracking throughout the night, but cannot execute altitude maneuvers. We have a solution to this cooking though!
Weather soundings play a part in making a forecast but it's not the only thing? Many times per day we collect weather station data, correlate against past forecasts, bias correct, etc.
Forecasts are made of so much input data it's insane, yes balloons matter but it's not the only thing. They are the only decent source of conditions aloft and the jet stream is main controller of our surface weather so it makes sense.
I'm amused at how upset people seem to be about being asked to sign up and pay for this service. I don't remember seeing comments like this on other launch threads on HN.
Sign up for free, try it for...not free. Go fuck yourself!
I’m glad you have the pricing clear in the front page so I won’t use it, unlike another similar service I tried that after wasting my time generating logos with prompts and details only to ask me to pay to download it..
My prediction. You won't survive a month with this kind of behaviour. ... Milking money out of people is outdated and obsolete approach.
After all the work I had sketching a penis I got hit with the paywall without even seeing a single result
So, I just logged in and there is no free plan at all to try your service as well as no way to delete my profile.
I went through the trouble of signing up, drew my sketch and prompt, clicked "generate" and it just sends me to a pricing page with minimum spend $20/month
Come on, just let us try the damn thing one or two times - nobody's actually going to be satisfied with the first logo they generate, so it's not like you lose the sale.
I think you need to give us the ability to test the creation of the logos before signing up. Maybe ask for payment later on when the user wants to download the logo?
I didn’t even get to see it work before being hit with a pay link.
I don’t really want to sign up before playing with it, maybe you can find another way to achieve your goal?
This is really cool. Could you build a few of these and make an interferometer?
I remember reading that interferometers are usually all connected by physical cables with physical loops to make sure the incoming data is combined at exactly the right time. But are we at a point now where that can somehow be done intelligently in software? Or are these little RTL-SDR's not accurate enough to even begin trying that?
Broadly speaking, the main prerequisite to interferometry is to make sure all the RF circuits are phase-coherent, meaning that all the oscillators are operating in lockstep within some tolerance.
The accuracy needs to be within about 1/10th of a period, give or take. I'm not sure the min/max frequency range for this system, but I saw 1.4 GHz in a screenshot, which would yield a tolerance of about 0.1 * (1 / 1.4 GHz) = 70 picoseconds.
That is achievable with the right hardware, but unfortunately the RTL-SDR doesn't include an option to use an external clock reference. As a result, each dongle's ADC sample timing and RF synthesizer phase will constantly be wandering relative to the others. It's not a one-time calibration; it's an ongoing random walk that changes on a millisecond-by-millisecond basis.
In theory you might be able to pull it off if you had a separate emitter in view of each antenna, calibrate each unit based on that signal, and then synchronize everything in software. But at some point it's easier to just use hardware that has an external clock input, and avoid the whole problem.
The always entertaining saveitforparts Youtuber tried to make a "very small array" from a bunch of used satellite mini-dishes. He ran into all sorts of fun issues with his ad hoc construction.
I admire the guy for having as much fun with his failures as his successes.
The lesson I've learned (I build scientific instruments as a hobby) is not to chase the latest, greatest ideas. Many of them only work because the implementors have access to a huge knowledge base, excellent parts and facilities, and really smart people to help debug the inevitable problems.
I focus more on maximizing what i can get out of a simple hardware setup, which means skipping anything that involves complex digital analysis or extremely sophisticated and sensitive equipment; it means more time having fun and less time debugging problems where I actually don't know enough to debug the problem.
In this case, the coherent source is the celestial object being observed. The problem here is that the "combining" step is being performed in software, and the sequence of digital samples have time- and phase-offsets that are ever-changing.
The two options are to keep those offsets under control (i.e., lock everything to a common clock) or to rapidly measure the offsets as they change and try to compensate in software (difficult).
In intensity interferometry the phase is not measured, and the timing accuracy is I believe only proportional to the desired effective bandwidth of the measurement. It was done in 1950's with bandwidths ~10 MHz -> 0.1microsec accuracy, should be do-able with SDR. Intensity interferometry is a bit of mind-twister...
(Another at first surprising thing is that radiation received from celestial sources is only coherent because of their very small apparent size -- the sources themselves are not coherent at all, because their physical size is very large)
Is there a better frequency available to amateur hardware that would give tolerance within more reasonable limits?
Without a shared external clock reference, i.e. over longer distances, how expensive does the hardware get if you want to be able to accurately measure the time/phase wandering to correct in software?
Just curious if it’s a limit of the low cost RTL-SDR or if it’s a harder problem than that (or both?).
Over local distances, all you need is a shared clock reference. As others have pointed out, the RTL-SDR dongle can do this with mods.
Over longer distances, this is an active area of research with many different approaches for various applications.
Two recent well-known examples are the "Event Horizon Telescope" (the network of radio telescopes that has been generating images of black holes) and optical frequency combs (a recent demo published in Nature achieved time-transfer accuracy of a few femtoseconds).
There's simpler options if all your equipment is in the same building, but yes, GPS-disciplined oscillators are a really great way to sync up clocks anywhere in the world to within a few nanoseconds. That said, it doesn't really help here because a) nanoseconds not picoseconds and b) the RTL-SDR doesn't have anywhere to plug it in.
No, it's not enough to know when samples from different SDRs were taken, they must all be taken within a very small interval (tens of picoseconds in OP's example). What the SDRs really need is an electrical signal called a "trigger" that tells them to read their sensors at that exact moment which is what GP meant by plugging them in. You can use a second GPS enable device to generate that trigger but synchronizing those triggers using time of flight is very hard.
Maybe you can bruteforce time correlation by shifting every source slightly, trying a yuge number of shift combinations and seeing which combination gives the 'best' (most correlated) output on a known signal? I've done that for many out-of-sync systems, and since I've started using GPUs in the two Os I've become obsessed with brute force methods :-)
Theoretically yes but using GPS time of flight to synchronize sensor triggers to within tens of picoseconds is so far outside of "amateur" that you might as well incorporate and start replying to DoD RFPs.
When your math starts requiring relativistic physics, it's a lot easier and cheaper to just run some fiber.
You'll want to look at the KrakenSDR. It's basically a circuit board integrating 5 RTL-SDR chips driven from a single clock source and calibrated for phase coherency, for the express purpose of doing radio interferometry amongst other things. They're in stock at Mouser.
Whoa yeah this looks like it would be easier than doing it all by hand. $466 at Mouser. Interesting!
Looking around on Google, I don’t see anyone that has tried using this for interferometry yet but the KrakenSDR team explicitly mentions that it can be used for interferometry.
They were pitching it for use in bistatic passive radar, which is a very similar application, until somebody put the fear of ITAR into them. Some of that code is probably still lying around.
We use images from istock.com (Getty Images) for the paid plans and unsplash.com for the free plan.
We ask gpt-4 for a good keyword that we can search on the istock.com API based on the business description in order to get images that fit the website out of the box.
Yeah! We made our own landing site using our own tool. That was pretty fun to do.
Asking the AI to edit things on the page is really fun. It's crazy what changes it can make with such little context.
For example the form where you input your business name and description. We just gave the AI a prompt of "turn this section into a form that posts to https://... with a business name and description" and it gave us exactly that, first try.
Hi! I’m Todd, the solopreneur founder of Prerender.io and I created that $1,000,000/year AWS bill. I sold Prerender.io to Saas.group in 2020 and the new team has done an incredible job growing and changing Prerender since I left.
$1M per year bill is a lot, but the Prerender back end is extremely write-heavy. It’s constantly loading URLs in Chrome in order to update the cached HTML so that the HTML is ready for sub-second serving to Google and other crawlers.
Being a solo founder with a profitable product that was growing organically every month, I really didn’t have the time to personally embark on a big server migration with a bunch of unknown risks (since I had never run any bare metal servers before). So the architecture was set early on and AWS allowed me the flexibility to continue to scale while I focused on the rest of business.
Just for a little more context on what was part of that $1M bill, I was running 1,000+ ec2 spot instances running Chrome browsers (phantomjs in the early days). I forget which instance type but I generally tried to scale horizontally with more smaller instance sizes for a few different reasons. Those servers, the rest of the infrastructure around rendering and saving all the HTML, and some data costs ended up being a little more than 50% the bill. Running websites through Chrome at scale is not cheap!
I had something like 20 Postgres databases on RDS used for different shards containing URL metadata, like last recache date. It was so write heavy that I had to really shard the databases. For a while I had one single shard and I eventually ran into the postgres transaction ID wraparound failure. That was not fun so I definitely over provisioned RDS shards in the future to prevent that from happening again. I think RDS costs were like 10%.
All of the HTML was stored in s3 and the number of GET requests wasn’t too crazy but being so write heavy on PUT requests for recaching HTML, with a decent sized chunk of data, the servers to serve customer requests, and data-our from our public endpoint, that was probably 30%.
There were a few other things like SQS for populating recache queues, elasticache, etc.
I never bought reserved instances and I figured the new team would go down that route but they blew me away with what they were able to do with bare metal servers. So kudos to the current Prerender team for doing such great work! Maybe that helps provide a little more context for the great comments I’m seeing here.
It looks like the hacker was able to log into the website's hosting provider account. The hacker then changed the DNS to point to a different server that had an exact copy of the blackwallet.co website but with a small injection of some javascript to siphon off the master keys that people were entering.
Sonnet 4 has definitely been the best model for our product's use case, but I'd be interested in trying Haiku 4 (or 4.1?) just due to the cost savings.
I'm surprised Anthropic hasn't mentioned anything about Haiku 4 yet since they released the other models.
reply