They want to be able to zoom in and zoom out as needed while analyzing aggregated user requests to better understand the different kinds of ways (well-resourced) actors use to distill their most capable models.
> "The data will help us defend against complex and novel attacks (including new jailbreaks and attacks that operate across many requests) as well as help us identify and reduce false positives."
> "Some attacks only become visible across multiple requests. Best-of-N jailbreaking, for example, sends hundreds of slight variations of a prompt in the hope that one will work. Larger patterns of misuse, such as state-sponsored espionage or data extortion campaigns, only surface when our safeguards classifiers can zoom out across many requests. Detecting these threats requires temporarily retaining prompts and outputs so they can be analyzed together, rather than one at a time."
Those studies have all, to my knowledge of them, shown no system mode that is better or worse than any other. The USA for many categories of care has often in those studies had worse wait times compared to nations with fully nationalized healthcare.
Heck when I went to college in Ohio's capital, the recommendation on how to get psychiatry or therapy as a new patient was to call the suicide helpline and claim that you were suicidal. That would get you a new patient appointment within 72 hours versus over 3 months on average for patients looking to get into care through normal channels.
Interesting you weren't able to find any examples. Here are some quick ones for CT scans in US Canada and UK.
The US maintains the highest hardware capacity of the three nations with roughly 43 scanners per million people and performs around 245 to 290 scans per 1,000 residents per year. non-emergency wait times are 1 to 7 days. The UK has 10 scanners per million people and ~100 scans per 1,000 residents, using centralized triage to keep non-emergency wait times between 1 and 6 weeks for NHS targets. Canada does 160 scans per 1,000 residents and 14 to 15 scanners per million, with wait times of 5 to 9 weeks.
When you consider that the US population is not fully covered by insurance, the number of scans is even higher
Okay, but that again seems to just be a few selected data points, rather than an actual trend, backed by peer reviewed study? Japan, Australia, and Iceland all have CT units per capita that exceed the US (~ 112, 70, and 50 /MM respectively), and all of three have universal care, two are full on single payer.
No, it's just all of the device models. Even the OSS tools take up almost 20 GB of space once you get all of the suppported device models which is not many compared to what Vivado supports.
Why not both? Look at the installables from Cadence, they are full of redundant garbage, broken rpaths, etc. Tools that don't even need to support devices have 5-8GB compressed installables with 12-20GB of on disk space being taken up.
Horribly broken installers and a mess of environment variables needed to get the software to even run.
As someone who actually worked for one of the largest FPGA consumers in the world, Xilinx didn't benefit at all from the community and hobbyist uptake. It was just an unnecessary expense and distraction that got in the way of their core mission. Around 5-10 companies make up around 70% of all FPGA silicon sales and the single largest end user is the U.S. Government and its customers (NATO members, Israel, Australia, Japan, South Korea, etc.). Now that cloud took off, FPGAs are somewhat more used in cloud deployments but usually as a stopgap measure until an ASIC can be developed. And putting the lines on the football field has finally stopped being one of the largest consumers because the broadcasters finally learned that they can load new programming onto the FPGAs instead of just ordering new hardware every time something changes.
> No. I said the low-end of FPGA sales is getting eaten by microcontrollers and the high-end of FPGAs sales is probably about to get eaten by custom ASICs.
You have absolutely no idea what an ASIC costs compared to a FPGA. A FPGA that can compete with a tinytapeout chip costs a few dollars at most in extremely low quantites. Something high performance would need probably TSMC 12nm or similar at a minimum. At that point, you're talking $1M+ between licensing fees and direct costs to just go on a shuttle. If you want to make your own higher volume run or can't wait for shuttle spot, you're looking easily $5-10M minimum for your first 6 wafers. Comparatively, FPGAs competitive with TSMC 12nm run from a few hundred dollars up to several thousand dollars each. So for low volume, they're very competitive.
Both SystemVerilog and VHDL have AMS extensions for simulating analog circuits. They work pretty well but you also pay a pretty penny for the simulator licenses for them.
I use AI for side projects because Google gives me a stupid large number of tokens that refresh every 6-24 hours on my existing $10/mo Google One plan. I see it as my civic duty to help increase their costs by producing slop that I generally throw away anyways because it doesn't actually work after it gets generated.
At work, I was told to use AI but it doesn't actually work for anything that I couldn't have handed off to a brand new undergraduate intern. So I use it for things that I don't care about then go spend twice as long rewriting what it output because it made the task longer by being wrong.
In my systems, I just go to an error log that gets posted to a Slack channel then go to the the log file and grep for full message that got dumped to Slack. That then gives me everything that happened before and a state dump after. That state dump can be given to a program to tell us if any state errored and what happened before tells us what the expectation was and what the precise error was. Using a LLM would just be slower and more expensive for this.
I can't get an LLM to properly handle analyzing a single 200K+ line log without making things up so whatever anyone is saying about this "working" is probably a lie.
