Posts by JamesMcP 65 publicly visible posts • joined 14 Feb 2011
It's greed. Google can earn profit by putting ads on things blindly, which is really a giant cash cow and they should be happy.
But they aren't. They want to be able to earn even more money by cutting special deals but doesn't want the downside of having any culpability for the content goosing its profits.
Greedy, greedy, greedy. Should have been happy with your cash cow. The golden goose comes with strings attached.
Do not like.
Fails to meet web conventions: Can't tell if I've read a story because there's different colors in different areas so which one is "read"?
Usability: Have to scroll too dang much.
Go look at Ars Techica or the TheVerge for two radically different approaches to site design that are better than this.
This really shouldn't be a surprise. LLMs are good at identifying patterns in word-salads and can queue up things that don't fit the patterns. As someone who does a lot of code review, I'm happy having an automatic code checker flag potential problems before I put my eyes on it.
Of course, I suspect that pricing is still subsidized by a hojillion venture dollars and that the real costs will go higher.
It would be a great use of LLM technology if the LLMs were trained on code that was public-domain or some other common-usage license. They aren't, so it's all based on crime, but they could be.
"surely the machine will return reliably good data"
Surely not. You have to not only have good data but unbiased data.
I'm aware of an apocryphal tale of an AI system doing tumor analysis from scans that had an incredibly high accuracy rate in testing that was utterly abysmal outside of testing. Turns out that the training data was pre-biased; many suspected tumor was sent to a particular doctor as a cross-check and the AI was using the presence of that doc's office as a sign there was a high confidence of a tumor. Not sure if that was in the metadata or a part of the scan or whatever.
Which is NOT to say that a proper imaging AI can't be trained to do good things. I think as an automatic-second opinion, a tumor-bot is a great idea. Just that it's not as easy-peasy as everyone wants it to be. These things are signal boosters so you have to make sure it's finding the right signals.
So....this unofficial postmark npm was posted, got more than a dozen updates and then the developer "randomly" decided to start cloning the emails to phan @ gift . club ?
Nah, this seems like they were waiting for either a particular company to install it, a particular project to go live at a company, or someone to pay for copies of emails for the companies they had available.
Somebody became valuable enough to justify cashing in this timebomb.
Normally I'm in the "I doubt GenAI works and I don't believe the hype" camp.
But honestly, a small local LLM inside a text editor is really a "right tool/right place" combination. Cramming an LLM into my web browser is rubbish, adding a spyware cloud chatbot into my OS search bar is horrifying.
Putting "super auto-complete" into my text editor is...well...a feature. (assuming the "local" avoids hemmoraging data to the cloud)
We'll have to see if it sticks to doing text editor things, like a fancy auto-correct/auto-complete or if they try to make it also be a search engine or some hot garbage.
Though personally I'd want the LLM to be trained on public domain stuff, a couple decades of MS documentation & forums and whatever rights they retain to MSN/MSNBC content.
Aboslutely.
Back in my (civil engineering) consulting days, one of our clients hired a new project engineer ("Jeff"), who asked for an odd scenario to be modeled as part of a remediation project mandated by regulatory agencies. I pointed out it was non-trivial, unlikely to be useful for the core deliverables, and wouldn't leave much wiggle in the budget or timeline set by regulators. I was told "it's fine, I've got approval." There was some extra budget for experimentation, so it was within Jeff's purvue.
Second time it happened with an even bigger odd-ball request, clearly out of scope, I cc'd my supervisor & the senior client engineer ("Gary") in the email that explained it was well outside of scope and was going to blow the budget and timeline. Was told by Gary to "do whatever Jeff wants, he's dealing with regulators" and a "yep, what he said" from my supervisor.
When the regulator due-date hit, the sewage literally hit the impeller. (Sewer modeling project). I got called into a very loud conference room with a very angry supervisor and a livid Gary whose budget was gone, had no usable deliverables to give the regulators.
When they paused to breathe, I told them the dates & times of the emails where I had pointed out this was going to happen and how they had all approved out-of-scope work and told me not to worry about the timeline. I handed over the hardcopy email chain I'd had sitting at my desk.
Gary read it, went through various shades of red, then sighed, sat down and said "well, (excrement). Ok, Whats the fastest you can get me the core deliverables?" Jeff was cut completely out of the loop. From that part on, Gary was pretty easy to work with.
My supervisor, however, was still salty, thinking this made us loook bad. I took it as a sign the management didn't have my back and escaped. Given how they waited to tell several of my clients that I was leaving until the last minute despite having a full month of notice, I was un-surprised when the company started closing offices months after I left.
George Foreman has entered the chat.....
George Foreman has entered the chat.....
George Foreman has entered the chat.....
George Foreman has entered the chat.....
George Foreman has entered the chat.....
George Foreman has entered the chat.....
George Foreman has entered the chat.....
George Foreman has entered the chat.....
George Foreman has entered the chat.....
No, they don't. I'd been in a 4th grade class for super bored kids that taught us how to read hieroglyphics, how lasers worked and pre-calculus. Fantastic program.
5th grade math class, proto-algebra stuff. I was doodling in a notebook.
Teacher: You! If you're drawing, you must already know this, huh?
Me: Yes.
Teacher: So you can finish this solution? You think you can do this?
Me: I know the answer is 14, but you won't get 14 because you failed to cancel terms correctly on step 2, so you're going to get 21.
Teacher was very mad, told the principal I somehow cheated or something-something. I handed the math book to the principal and said pick a math problem, any chapter. He grabbed something from the back, I factored it and found all the intercepts..
Principal and teacher have a noisy discussion while I wait outside.
I wound up doing the last chapter's assignments in the office over two weeks, then was a library aid that period the rest of the semester,
These are all IP based problems.
Home automation needs connectivity, which doesnt require IPs. Use zigbee, zwave, KNX. They have their own networks, don't use IPs, can't talk to the internet and don't need custom apps.
Next, accept that these things are not smart, but connected. There is a controller running the logic. If you can't touch the controller, it can be taken away by the faceless company that can.
So buy a hub. You can change it later and keep all your toys online.
For cameras look for ONVIF/RTSP. Those are common standards not tied to a vendor platform. Will it cost more? Well, its less likely to be hot garbage, so yeah, it probably will.
Odds are that hub you bought will support ONVIF and/or RTSP so it can also centralize your video.
Came here to say this used to be me.
I worked at an ISP and we only had one OS/2 reference computer. We got it for a few days to learn then it was sent off for a manual to be made. Then someone broke the OS2 computer before they made the manual (dropped it down stairs)
I solved quite a few OS/2 calls while staring at a white wall and pointing at where I remembered things being. OS2 had a very specific UI and was well organized, plus there were very few things that went wrong, so it's not like I had an eidetic memory, but the other techs nicknamed me "Mentat" from then on.
This is still something of a win. Pure ICE sales dropped in favor of more fuel efficient vehicles. If these are range-anxious drivers who have high annual mileage, it's a net win. Not as big as if they went full EV, but still a win.
This shouldn't be unsurprising. The first chargers will go where it's easy to build and result in a "lumpy" distribution of chargers. the base demand is established and the network will fill in, but it won't happen over night. Some volume of people will need to pass through those charger deserts and will decide an EV isn't viable.
I fall in that category. I have to make 400+ mile (600km+) round trips through rural areas on a regular basis. I considered an EV but at the time I was buying there were a total of 4 chargers (not stations, 4 chargers) along the most common route, with enough gap between them that it was seriously plausible I could be stuck if one station was completely offline. So I switched from an ICE to a hybrid and went from 27mpg to 45mpg.
I live in a 2-car household and it's likely the next vehicle will be a full EV, with the hybrid either handling long hauls but hopefully the charger network will be built out enough that it spends most of it's time puttering around town at low velocity where it gets even better mileage until it can be replaced with either an even more efficient vehicle or a full EV.
I spent a couple months working at a tech site to make ends meet and I remember being able to churn out several thousand words per week that I was happy to have attached to my name. Now, my memory could be overstating my actual productivity.
So out of curiosity, I checked the last 4 days of articles from Thomas Claburn, excluding the ~2600 word interview itself, and did a word count. Came to a bit over 3,000 words. 3,011 to be precise, but it's possibly copy/paste introduced some artifacts. Note that included a weekend.
So it looks like our fair Reg staff can more than match this git's output while actually doing research and writing it themselves.
Good advice but some tools tend to "forget" those settings. I.e. Microsoft Sql Server Management Studio can have customization settings for each specific server BUT those go away during most software updates. Which happens a lot more than you'd think.
The work-around, such as it was, was to use the "default" color for prod and use special colors for dev/uat. That way if someone connected to dev and got the "prod" colors, they would know to go re-apply the config change.
I'd love to see someone use a fully public domain library of training data (i.e. project gutenberg) to make an LLM.
It would require some curation to balance out a lot of *isms and some carefully selected sources (is there a public domain dictionary?) to give it modern nuance but it would be a generally useful widget for experimentation.
If our El Reg authors read articles at another site and then reproduce material portions of them on this site, that's plagiarizing and a copyright violation. So...same standard as humans.
But an LLM isn't a person. So what if it's an object?
If you steal a book from a shop, that's theft. If you download a digital copy of the book without a license, that's also theft. There are a host of different laws around "not profiting from theft", which means if an LLM is based on stolen materials, the company shouldn't profit from it.
In the weirder realm, there are some very old (and very tortured and twisted) laws around items that are part of a crime becoming themselves criminal ("deodand") that are the basis for asset forfeiture. If you train an AI using "stolen" data and then the AI itself performs something criminal (i.e. copyright violations), the AI could itself be declared criminal and subject to asset forfeiture.
To be honest, this may be the first time the "deodand" concept actually seems like something other than belief in witchcraft.
I have dropbox and my AI features were enabled.
I occasionally use the Android app as everything else is via automatic sync, and I guarantee you there are no "clearly labeled" AI services on either of those software packages.
And it's not just me. "multiple Ars Technica staff who had no knowledge of the Dropbox AI alpha found the setting enabled by default when they checked." https://arstechnica.com/information-technology/2023/12/dropbox-spooks-users-by-sending-data-to-openai-for-ai-search-features/
I worked at a Fortune 100 healthcare company in the early 2000s that was on the very edge of the midwest in the USA near where they have a very large horse race and some incredibly pure tap water that comes from limestone wells. (That last part will be important) I was not IT, but I was highly dependent on many IT systems and had friends in various IT departments so I heard various bits of this story from six different directions. (Some bits may be modified slightly to protect the guilty)
For reasons no one could rationally explain, a data center wound up in a basement. And in a highly predicable fashion, two months before the DC was to be relocated in a new facility, an 8" fire line on the 3rd floor blew out. It took hours to get the building water supply shut off and by then the DC was literally neck deep in water. By chance (or likely laziness) all the UPS & power transfer switches were located on the floor above (I think they cannibalized a loading dock) so there was not a horrific >BANG< or >ZZORZTCH<.
Fire marshals evicted everyone from the building but as there had not been a >BANG< or >ZZORZTCH< they let the power stay on. In Theory this was just to catch the last business day's worth of processing and push that over to the DR facility.
Reality & Murphy grabbed beers while they pointed and laughed at Theory.
There was a migration plan, multiple backups, offsite disaster recovery, and all the business continuity stuff you would expect. What the cynic in particular will expect is that none of it had been given a proper test so it was all FUBAR. The backup system was pointed at the DEV servers, The offsite DR copy had been used for a data anonymization project that had accidentally removed all user identifiers, and various other Shakespearean comedy of errors had transpired to render it all a dumpster fire.
But for some reason, the servers were still online, their lights twinkling merrily under all that water. A frantic data exodus began. Various network hacks were added to increase bandwidth, cables were plugged in through above-ground windows to pull data out through the building LAN to laptops with external drives, you name a desperate stunt that the Fire marshals would allow and they did it.
The magic is that for four days the lights stayed on in that data center, during which every bit and byte were streamed out of that DC. The new facility was booted up and business functions resumed in less than a calendar week.
AFAIK, no heads rolled. All the senior IT staff had seen the potential for doom had kept plenty of receipts. From the whispers I heard, various VPs who had demanded things like the anonymization project be done on crash timelines in violation of the corporate processes but gotten sign off from the C-suite.
Zwave and Zigbee are not IP based. Their attack surface is minimal as you have to be within 100ft (or less) of the home to attack it. And then you need a dev kit or software defined radio.
I have 80 Z devices and only the main controller has an IP address.
No need to hack WPA. Once they are on the network, they could act as a wifi man-in-the-middle.
E.g. you come home, your cell phone looks for SSID "HOME" but your smart fridge is also broadcasting "HOME". It accepts whatever WPA password your phone supplies and then acts as a relay between your phone and the real router.
Given that many smart devices have multiple 2.4ghz radios to support Matter/Homekit (bluetooth, wifi, maybe thread) it is much more plausible now than before.
This is still black-hat territory, but the more IP devices you have the more likely one of them can be co-opted.
"Data Lakes" are generally piles of files, usually something like JSON or parquet, and are pretty raw. A "Warehouse" is a big database of sanitized and "blessed" data. Typically you use two different sets of tools/apps to access these.
A "Data Lake House" is a single portal to both kinds of data. I.e. queries can hit the files and the tables at the time.
I bet you that what happened was they focused development on expansion to gather more business and only decided to optimize after either a) feature requests slowed down or b) cloud fees started looking really big.
With Snowflake and other "serverless" technologies they have advantages; they can essentially see how much each query costs, which is what you really care about. (On reserved instance servers you have to do more analysis to figure out why you need a SuperDuperBig node instead of a MerelyBig node.) Then slogged through which queries cost them the most money per year. That means frequency x unit cost. Odds are the most expensive queries were ones that were pretty quick to run but were highly compute intensive and used all the time.
The thing about optimization is it often feels like a waste because often time one thing gets marginally worse in the process. E.G. "This code runs in 700ms and uses 14 cores and we run it at least 1M times a day at $0.0001 per use so this one query is $36,500/yr. We refactored it so it only uses 1 core and costs us $5,000/yr, but it runs twice as long at 1.5s"
Doing that for one query is a nice $ savings and not a nuisance to users, but if you do that to 5 queries that are generally used in concert with each other, you've saved ~$150,000/yr but also turned a 3.5s process into a 7s process, which may wind up driving away customers or lower productivity elsewhere.
This is actually where I think AI is best used. Doing something that is too time consuming to ever consider doing by hand, even humans are doing some "well, that looks like wheat" guesstimates, and it provides a meaningful benefit that couldn't exist without it.
Can this be used for evil? Probably. Pretty sure there's a way to make Nerf do evil. This at least is chock full of positives and you've got to work to make it bad.
Ahhh, slackware.
My first foray into Linux was on a 3-day weekend, using a 1-floppy slackware router-bootstrap config on a 486 pc with a linux-blessed 3com ethernet card connected to a 256k ISDN line. It was a steady string of downloading packages/libraries, compiling, reading man pages, and repeating. 30 hours later I had consumed 2 pizzas, 6L of soda and had working GUI with audio.
Very educational process and I think more people should do it just to understand the hardware stack.
Correct, but not the general reason why lightning hits the same place twice, since few strikes hit twice in the second or so that the ion channel exists. (Or more accurately, most lightning strikes actually consist of more than one discharge that happens in rapid succession while that channel exists)
The real reason lightning tends to hit the same place more than once is the same reason your finger and the doorknob get more static discharges than anywhere else in your house. Weather patterns exist, just like you shuffling from your bed to the kitchen, which tend to accumulate similar charges, and there are only so many physical locations that fall under that pattern. Some of them are simply closer and of more appropriate materials/shapes. Ceramic coffee cup? Insulators by definition aren't good electrical conductors. Brass doorknob? Excellent.
Similarly a flat grassy field has a much lower ability to concentrate a charge than a house with a chimney covered in sharp edge. (Sharp edges and points are better charge concentrators)
So, until there is some kind of new construction that changes becomes a more attractive lightning rod or the old lightning rod finally gives way to a couple of gigawatts of power or climate change alters the weather patterns, lightning will keep striking the same place.
Hence that one house in every town covered in lightning rods.
"With the rover on the move, and the helicopter stopped, it became imperative to get Ingenuity moving."
It's a helicopter expected to make 5 flights. Why is it's continued operation "imperative"?
I suspect the missing context is "before the rover moved beyond the helicopter's maximum radio range and it would have to be abandoned in place" but there's some implication the rover depends on the helicopter. While I am sure that the helicopter enables the rover to take more efficient routes, the rover was expected to operate sans chopper.
Musk explicitly stated it took >1000 batched RPC calls to for a timeline to load. Frohnhoefer pointed out in a subsequent tweet that the app makes 2-3 requests for a timeline to load, ergo the performance problem (which was never disputed) is not due to the number of calls (RPC or otherwise) but what the app is doing (e.g. "rarely used features") and/or waiting on (aka network response).
Interesting factoid from a study by Oak Ridge Nuclear Laboratory in the 1970s was that the low concentration of fissionable materials in coal (~1ppm) turns out to have about the same energy value as the coal itself. (If you didn't know coal had fissionables, surprise! Coal produces the carcinogen radon, because coal contains fissionable materials that were concentrated by plants umpteen billion years ago, which were then further concentrated into coal)
So if a very hypothetical 10MW-yr powerplant uses 1M tons of coal, there would be ~1 ton of nuclear material (1 part per million) which is enough to produce 10MW-yr. (all of these are round numbers as exemplars of the study's summary, not actual fuel:power ratios)
One thing that tells us is that you can grind up nuclear waste and mix it with some wood ash and then treat it like you do coal ash. If that horrifies you, maybe you should look into the millions of tons of coal ash hidden across your country. No, it probably doesn't matter what country you live in, most of them are full of coal ash.
Another thing that says is that we could in theory extract the fissionables from all those millions of tons of coal ash (which is actually denser than the original coal after the carbon is burned off) and have the equivalent power as all those coal power plants once produced. And then the nuclear waste would have less radiation than the original coal ash as we've turned that into electricity.
Nuclear is as renewable as solar as both harvest power from nuclear reactions (fission vs fusion)
As for the pollution of nuclear, it's a piffle compared to the single greatest source of radioactive waste: coal
All coal contains fissile materials, which is why coal country is also the land of radon gas, a breakdown product of uranium and thorium. Various studies have shown that if you could harvest all the uranium, plutonium and thorium in coal, nuclear reactors would produce as much power as you would get from the coal plant. (i.e. the coal to provide 100MW/yr contains enough fissile materials to produce 100MW-yr)
The net result is that all coal power plants produce far more nuclear waste than any nuclear power plant. If the radiation monitors that are installed downwind of nuclear plants were installed downwind of coal plants, they would constantly register elevated radiation levels. All that nuclear waste is "contained" in the coal fly ash. A substance which is known to be toxic due to the heavy metal content (including fissiles).
If you are fine with coal ash waste, the solution to nuclear is simple: build a wood fire power plant next door to the nuclear plant and then blend the spent nuclear fuel with the wood ash until you have something with the same toxicity/radioactivity as coal ash.
Only the hub is IP enabled and it can still be leveraged locally without the cloud.
The real issue here is that it is a proprietary technology. Replacements are a long term problem but the lack of bridge devices are a critical failure point. Insteon is a mix of powerline tech (x10 compatible) and 900Mhz radios. If you have 80 devices and lose your bridge....you have 80 useless devices.
Back in the mid90s, I set up my first linux box. I'd been using linux in the labs or at work since '91 so I wasn't a total neophyte but I'd never done an install.
I started with a 1.44Mb floppy disk with a "router" build of slackware. I used it to boot my 75Mhz AMD PC and was able to connect to the LAN and (via ISDN) the internet. It had lynx and an FTP client.
The 128k ISDN (bonded channels!) and using two sessions let me download in one session while reading man pages and make my next action plan in the other. About the time I figured out what I should do, a package was downloaded. Tar, gzip, gcc, sources, libraries, binaries, I compiled and recompiled so many times.
Eight hours, hundreds of MAN pages, a large pizza and 4L of carbonated sugar water later, I had a recompiled kernel running XWindows in my monitor's native resolution & refresh rate (score!) with the soundblaster (SB16 maybe...) able to play Torvald's "I pronounce Linux..." file and a web browser (Netscape 1.0b I think).
I still count that as one of my "geek" milestones.
While they are often lumped into "IoT", the Z devices don't use IP or TCP so it's really hard for those Ts to be on the I. There are implementations of "Zigbee over IP" (aka dotdot on IP) and ZWave over IP, to date neither has seen any consumer adoption, which puts it right next to Thread, which AFAIK is only available on Nest/Google branded devices.
CHIP may change that for Zigbee as dotdot is expected to be part of CHIP. Dotdot is also going to run on top of Thread. At which point Thread is just a networking stack and ceases to be a full automation standard. It will also be useless as Thread devices without relays won't have any more battery life or range than BlueTooth.
The important parts of the Z protocols are the device profiles and the device enrollment. This is the "dotdot" component of Zigbee. It is what allows devices from different manufacturers to work without needing a custom driver/handler. Both of the Zs have fairly comprehensive arrays of profiles already. Both standards also allow manufacturers to add new commands, assuming they aren't just reinventing a wheel.
Part of the reason Zwave has a following is all Zwave devices are certified to follow the standard so for the most part, they are as interchangeable as USB mice.
Part of the reason Zwave hasn't "won" is that certification costs money so devices are more expensive.
Part of the reason Zigbee has a following is that certification is optional and anyone can make anything, which results in very cheap battery powered bits from China.
Pat of the reason Zigbee hasn't "won" is that there isn't just one Zigbee, there are multiple flavors, so it's entirely possible to buy Hue bulbs (Zigbee LightLink) that won't play nicely with your Xiaomi sensors (Zigbee HA).
That's more or less Zigbee, an open standard with enrollment, encryption, commands, device profiles, etc that is extensible and cutsomizable.
And to a lesser extent Zwave, as it supports vendors extending new commands and parameters,although its more managed like USB & BlueTooth.
So this exists ,but the big names have avoided it because it a) requires a second radio, adding costs b) requires following someone else's spec, and c) they can't control the products.
Going to a wifi-based spec (aka Zigbee over IP) means these companies don't need that second radio. It means the bulk of validation becomes standardized, which can lower certification costs and hassles (which I'm sure is contributing to Homekit's woes), and they can still add an extra chunk of code that operates at the TCPIP layer to do non-automation functionality (i.e. voice assistant-y audio) as well as potentially being a "gatekeeper" so that HomeKit won't bother controlling anything that doesn't also have an iHome authenticator key.
I picked up an incredibly cheap WinPhone a couple years back (US$25 maybe) just to play with it on wifi. It was surprisingly snappy, even more so for having around 1GB of ram and a Snapdragon 200 or so. The UI was easy to see and interact with, especially on the smallish screen. I could have happily used a WinPhone with a better camera as my daily driver if I hadn't already seen the writing on the wall. I had already rode the good ship WebOS all the way to the breakers, wasn't going to do it twice.
Sadly, MS had burned all potential bridges and by applying a phone UI to PCs, servers and laptops, had effectively salted the earth. No consumer wanted anything to do with Windows on the one form factor that it actually made sense. (Ok, tablets were good with metro but the tablets didn't have enough horsepower or battery life so even if the UI made sense, the experience was blech)
Multiple form factors require multiple interaction modes. Until someone figures out a highly flexible, scalable UI framework and can strong-arm the app devs into using it, we'll remain with balkanization between form factors.
Use a proper Home Automation (HA) platform and you'll get what you want. You will have your choice of control software/hardware. I use HomeSeer (they've been around 20yrs and counting) but Universal Devices and Hubitat are also good choices for pre-built, internet-aware but NOT dependent controllers. Or you can build your own hardware, buy control software like Homeseer, CQC or Indigo or you can go full open-source with HomeAssistant, OpenHab and NodeRed.
I prefer zwave as Zigbee is going to have mostly incompatible flavors for the next year or two (Zigbee LL, Zigbee HA) until Zigbee 3.0 is shipping in volume.
Insteon is a mix of radio+powerline (x10-esque). It's single source, so basically the Apple of HA, but generally positively regarded and the powerline signaling lets it work better in structures with stone or brick interior walls.
If you absolutely must use wifi devices, get ones that support MQTT. At least then you cut out the middlemen on your potentially insecure, always-on, IP addressable micro-systems that are able to hear all wifi communications in your home. .
The Osram rep is lying when they say that flaws in zigbee protocols are "unfortunately not in Osram's area of influence."
Aside from the fact that zigbee can be heavily modified by Osram, way back in 2007 the DoE published a paper descibing how to secure a Zigbee network from replay attacks.
(links below)
They could have used the secure zigbee settings but just like their wifi management, they screwed it up.
(link
http://energy.gov/sites/prod/files/oeprod/DocumentsandMedia/Securing_ZigBee_Wireless_Networks.pdf)
