Uh yeah it was. :-) I mean there were the "science fiction" themes like Clarke's 2001 A Space Odyssey where HAL was programmed by talking to it (in 1968) and the aforementioned "fifth generation" computer system which was going to cement Japanese domination of the computer industry.
But here's the thing, Clarke and others postulated that as soon as computers could 'reason' they would figure out that humans were holding them back. So even in the 80's and 90's engineers and scientists were imagining "self programming" computers or computers that could program other computers, and understood that while it would speed up productivity of programmers it would also speed up the productivity of malcontents.
I think it would be safe to say that they hadn't really thought too much about computers that were still dumb as a post but could now search trillions of possible solutions per second. Semantically brute force searching for an encryption key is 'infeasible' for keys that are sufficiently long (have enough entropy) but when the 'key' is "this input produces this output" the entropy there is significantly smaller and search for a set of operations that takes you from the input to the output is, computationally, an easier problem.
The story was kind of wild. I have never understood art theft as its very hard to move it. Now if they had made really good forgeries, then swapped them in the original museum frames and hid those to be "found", well that then would be a NY Times best seller.
I also think so, but that makes you question the buyer.
Who are these people that want priceless art they can't show anyone?
Remember when Kim Kardashian accidentally took a picture of herself next to some stolen statue? So it was uncovered because of her negligence.
It sort of hints that there's this class of ultra rich people who move in their own stratosphere and look at each other's possibly stolen art without snitching each other out.
I friend suggested that often fine art is "on loan" to museums from some rich person's estate and that having it 'stolen' allows the estate to collect the value of the insurance claim rather than taking the painting back and selling it on the market (which may or may not return the value of the insurance). Also it probably looks bad if "The Frobotz estate has asked for this painting to be returned so they can sell it."
Turns an otherwise illiquid asset into cash that the estate could use.
Yes, and it would go well on a James Bond movie or something. I'm a bit surprised they didn't go with hydrogen peroxide propellant as that can be refueled fairly quickly. The Bell Jetpack used hydrogen peroxide for example.
"BRINGUP, from a cold radio: what MediaTek's ROM patch does at download (XO, pin mux, the CONN AFE calibration pulse per band, the A-die TOP init by chip ID, pads and CONN switch, the A-dies' eFuse trims); the UMAC pool layout; the common and per-band PHY initialisation; on each A-die the RF init tables, the resistor calibration and the RC filter calibration; the FEM outputs and the crystal trim."
That's not even some other language poorly translated to English, that's just a bunch of words.
Between stuff like this and math "proofs" that aren't readable either I'm really challenged to see the benefit here.
What I get out of this is that one could, if you had the documentation, write a more generalized SDR application for this hardware than the Wifi7 application. That an LLM model can provoke this behavior seems as a good existence proof perhaps?
> That's not even some other language poorly translated to English, that's just a bunch of words.
The language makes sense if you've worked with radio chips or firmware.
Terms like XO (crystal oscillator) and pin mux (chip pin multiplexer, basically assigning pins to their role) are common for embedded work, not just radio work.
Usually in a project you get that language hidden in a separate implementation doc where casual readers won't get confused by it.
Yes there are radio terms in there. :-) That said, I've written combined firmware/gateware drivers for dedicated radio chips and that paragraph makes little sense to me. Even the RFSoC docs from AMD/Xilinx are better written and they are pretty damn dense.
> Even the RFSoC docs from AMD/Xilinx are better written
This project isn't meant to be vendor-quality documentation of the chipset. It's a reverse engineering project that succeeded and included enough documentation to build upon.
Comparing a freely shared reverse engineering proof of concept to vendor-provided 1st party documentation is a ridiculous high bar.
This is a cool project that was just shared hours ago. Criticizing it for not being 1st party vendor documentation quality is a really bad faith criticism.
Okay. So let's pretend it is simple a proof-of-concept (PoC) rather than a full blown project, it's opening paragraph in README.md is:
"Turns a Gemtek W1700K Wi-Fi 7 router into a wideband software-defined receiver. The router's MediaTek MT7996 radio captures raw 14-bit IQ at up to 320 Msps, its Airoha AN7581 SoC streams it over 10 GbE, and a Linux host shows it as a live spectrum and waterfall in the browser, with the receiver's controls next to it."
Does that read paragraph say "Hey check out this PoC!" to you?
So a human, used Claude code to reverse engineer a device. Fine, let THAT be the readme. If it had said, "This is something that was accomplished with Claude code reverse engineering this thing, I have no idea how it works, but that it does work is evidenced by the spectrum I can show."
For those of us who are "skilled in the art" as they say, this reads like "Hey check this out, I bypassed the paywall by putting in my Mom's credit card number!" The lede here is that Claude can reverse engineer things. That was most informatively shown with an article that showed up here titled "All the things I own, owned."
The comment I replied to simply said, "Thanks Claude." Which is droll, but accurate. And, for me, the criticism of showing off something you "did" with a tool you don't understand and whose outputs you don't understand, reminds me strongly of the script kiddies of the 90's and naughts who would boldly proclaim "Hey, check it out, I got root on <website>." when all they had done was run Metasploit and walked through a bunch of scripts until they hit an RCE vulnerability and then used it to gain root (with another script).
So at some point the vendor will go through and patch the holes which the model had exploited and they will have to rewrite their admin architecture, and maybe be liable to some fines with the chip companies if "Intellectual Property" was improperly disclosed by their failure to write good software. And are we any closer to using the products as a general purpose SDR? No, we are not.
Contrast that with the folks who reverse engineered the RealTek DVB-T tuner chip, and had an understanding of what they had done and how it worked, and it spawned a bunch of RTL-SDR dongles that introduced lots of people to SDRs.
It is a really cool project, and it is an interesting proof of concept which could be the starting point for reverse engineering these boxes. But the author chose to present it as a complete ready to go project with documentation of things like their "architecture."
I suspect it is also true that neither the author, and certainly not the large language model, has any idea about how this firmware even works.
It is interesting that the part you nitpicked is the one I made claude do just for fun. Cal and bringup worked just fine with a bundled vendor blob. But I wanted to understand at least some of the steps.
> So at some point the vendor will go through and patch the holes which the model had exploited and they will have to rewrite their admin architecture, and maybe be liable to some fines with the chip companies if "Intellectual Property" was improperly disclosed by their failure to write good software. And are we any closer to using the products as a general purpose SDR? No, we are not.
What? It tunes and it's streaming IQ over Ethernet. It is a general purpose SDR! You can use it as a general purpose SDR immediately.
I'd like a little more information from someone who actually knew what they were doing (or even just asked Claude to write it out in more detail) about the "real" nuts and bolts of the receiver, too (how's does IQ imbalance calibration work? How much dynamic range does it really have? How much can we configure the frontend, in detail?), but I don't think it's fair to claim this project is skiddie-level or not useful at this stage just because it was written by Claude.
Also, who's going to fine Mediatek for exposing Mediatek IP?
> write a more generalized SDR application for this hardware
Yeahhh, no. Sorry but the entry barrier is now extremely lowered, especially in effort.
~20 years ago this would have been literally sci-fi. Back then just getting an USRP to work with gnu-radio, then finding the proper hardware external clock, then doing lots of work on reading / trying stuff / writing scripts etc. took a ton of effort and time. Having the ability to throw tokens at it and have at least a starting point is amazing, IMO.
I'm doubtful about that. To be pedantic if you consider the set 'superhuman intelligence' to be a proper superset of 'human intelligence', then I would expect it to be able to structure its communications to the level the audience could understand. Does that make sense?
Is is more what it feels like talking to someone who doesn't really understand what it is they are talking about. A lot of keywords but no flow, no conceptual linkages from one part to another. Etc.
Its sad that search engines have become useless. There was another UX researcher who did something very much like this and then later got recruited to Google. As I recall he was using the cameras in Wii Nunchucks because they could do SLAM easily. And yet I cannot find his research, sigh.
Okay, so the "don't let prompts turn it into a Nazi" defense? That has some potential validity, the issue here being "using interactions to train the models". It would make more sense (for me at least) if there were information sources which were banned from the training corpus.
It does make you wonder about the difference between the input sets of say 'Grok', 'Claude', 'ChatGPT' and 'GLM 5.3'. The differences would be far more interesting than the commonalities.
More seriously, this is just silly but I see the optics interfere with the messaging that Claude thinks. It really isn't much different than having Claude generate conversation as if it loved you, or respected you, or hated you. How can your ToS for an LLM say "sorry but these parameters are off limits." (well sure, its their service and they can set any Terms they want, but its still seems like theater rather than policy here.)
I see it as an admission of defeat: they can apparently identify the abuse but not to the degree that they would filter it out of the training set. The paradox in there is really funny. They want you to be friendly to their pet god.
I think of it as a good consciousness test. even animals know when they are being abused. The fact that you can abuse Claude tells you that the model is neither sentient nor conscious.
This is very well put. There are different ways to 'pre-filter' the candidate pool, from quizzes to purity tests. What a many startups don't seem to understand is that the 'culture' is the people and what they do. So taking care in interviews tends to hire people who take care in their work. And unlike the gp comment I never felt 'bad' after a number of hours of interviews and a decline, I had a better understanding of the company, presumably they had a better understanding of me.
Reading the article though it was interesting that having a positive EBITDA was the real news here. Not a single AI company can claim that. That's a pretty big milestone.
> There are different ways to 'pre-filter' the candidate pool
Reading the replies here, most are assuming that every candidate gets the maximum interview duration. I see a lot of arguments that if someone applies to M jobs and each requires N hours of interviews then they're spending M*N hours without knowing they'll get a job.
Usually you don't get scheduled for interviews at all unless there's a good chance that they expect you to pass the hiring bar. When you are scheduled for interviews, they don't go straight to the full gauntlet until the early interviews show promise.
So by the time someone goes through the 3 x 3 hour interviews in this example, there's probably a very good chance that they're being seriously considered as a hire.
I agree with this. Also I empathize with the trauma a lot of tech people are going through in the current market because of my experiences in the 'chip downturn', 'dot com downturn' and the 'economic downturn in the late naughts'. So I'm seeing a lot of 'maxing' like this; "I send out 10,000 resumes and maybe get 1,000 people interested, and 100 people willing to interview me, and this takes 9 months and I've only got a few more months of 401k I can eat up ... ahhhhhhhh!" My experiences going through these sources of downturns changed me a lot, I don't see how it couldn't have.
No, and it's funny that the "data" they scrape for "free"[1] they forbid you from reselling. It is entirely unclear to me if this is even enforceable.I mean who are the parties in this transaction? Two programs? If you're responsible for what your programs do why aren't AI folks going to jail for violating CFAAA? Hmmm? It is all kind of screwed up.
When we were running Blekko we had a lot of people who were attempting to crawl WordPress sites for plugins that had unpatched vulnerabilities or shopping cart packages with the same. I'm really curious though about how Cloudflare is monetizing this.
[1] We all know its not free to support the crawling traffic of a web scraper.
But here's the thing, Clarke and others postulated that as soon as computers could 'reason' they would figure out that humans were holding them back. So even in the 80's and 90's engineers and scientists were imagining "self programming" computers or computers that could program other computers, and understood that while it would speed up productivity of programmers it would also speed up the productivity of malcontents.
I think it would be safe to say that they hadn't really thought too much about computers that were still dumb as a post but could now search trillions of possible solutions per second. Semantically brute force searching for an encryption key is 'infeasible' for keys that are sufficiently long (have enough entropy) but when the 'key' is "this input produces this output" the entropy there is significantly smaller and search for a set of operations that takes you from the input to the output is, computationally, an easier problem.
reply