gok

Computer security is not a serious field. There is no other group that honestly feels "do what I meant, not what I said" is a sign of someone else's bug.

It's hundreds of milliseconds to do a a million iterations. A single time check is hundreds of nanoseconds.

So these guys are just dumping confidential tax documents onto OpenAI's servers huh.

see also https://news.ycombinator.com/item?id=39485940

Swiss trains from Stadler are in fact selling rather well in the US recently (Caltrain, e-BART, New Jersey Transit, Austin, Dallas, and most recently Metra in Chicago)

The car strength requirements only apply to mainline rail, not rapid transit like BART which runs on dedicated rails. And even on mainline rail, the FRA now accepts Europeans rolling stock as of a few years ago.

M3 Pro gets to 36GB, the M3 Max can do 128GB

The EU would surely fine them for that too.

It's a weird move by the FTC.

Kroger-Albertsons together would have less than half the marketshare of Walmart. It's closer to a third if you include Sam's Club. In practice it would be even less than this, because they're planning to spin off a quarter of the stores.

The decision to set the precedent that union shops can't buy each other is also odd. If non-union stores are allowed to expand but union chains are not, that is invariably setting grocery store unions on a path to shrink even further.

The common form of this bug is way older than 10 years and isn't really specific to macOS.

The problem is that audio devices do not have an explicit concept of "balance". It's set by adjusting the level of each channel. That volume adjusting process is not reliable. It often happens over an unreliable interface where some volume adjustments are just dropped. This hits a second problem: the audio device often has the ability to adjust its levels out of band with the host audio stack. Thus your laptop has to periodically poll the device to find out what the current actual audio levels are. When your laptop sees that the volume per channel is not equal, it doesn't know if this was caused by the device itself adjusting the balance or the volume adjustment message being ignored.

So for example:

1. User tells laptop to increase volume from 50% to 60%

2. Laptop tells device's right channel to go to 60% (success)

3. Laptop tells device's left channel to go to 60% (failure)

4. Laptop checks with audio device, sees 60% on right and 50% on left

5. Laptop changes audio balance to be nudged to the right

There are some ways to mitigate the problem. The audio stack could try to detect when its level commands are being dropped, but that risks the audio stack and device "fighting" over the correct volume. Better closer-to-realtime scheduling can reduce the risk of messages being dropped.

> Under these assumptions, Laura uses a computational technique called Dynamic Programming to evaluate all possible scenarios in the game tree. Once you factor in this more complex set of possibilities, the “62.5%” mirage disappears. Laura finds that if you are down by 2 TDs, G42D8 increases your probability of winning by less than 1%.

This is misreading the original source [1]. The <1% figure applies when you have not already gotten the first TD.

[1] https://punkrockor.com/2019/11/15/when-should-a-football-tea...

Because Tencent owns both Epic and QQ

Mostly small escort carriers which displaced around 10,000 tons. Smaller than some modern superyachts.

Do you think the point of vending machines is to equitably distribute soda?

Free/cheap parking is a driving subsidy. In fact the old prices were even worse than that; the prices were not high enough to actually keep spaces available, so it ended up being more like a lottery system. The correct solution would have been to just raise prices and fix enforcement, but the city dug themselves into such a financial and political hole that it wasn't feasible to do without the sale.

Yes, and they refused to appropriately increase pricing because it was a cheap political gimmick.

The city itself would be losing that revenue if it retained control, and even with the deal they still need to budget the loss in parking ticket revenue.

The primary reason the parking deal never goes away is that it allows the city to blame "investors" for doing something it was already wants to do (keep lots of street parking).

What people miss about this deal is that the city was literally losing money on parking meters before the deal because it was so incompetently managed. Chicago retains all the revenue from tickets, which are much more frequent now, so overall the city is making a lot more money from street parking today than it used to.

Now, ideally the meters would have been run correctly with municipal ownership, but that was apparently not within their capabilities.

Unfortunately the team behind all this was disbanded, and now Samsung just uses CPU uarches from Arm.

If you'd like to feel extra uncomfortable next time you get on an airplane, read a bit more about limitations of TCAS that still exist today. For example, it doesn't actually model aircraft performance, so TCAS occasionally issues resolution advisories that are not physically possible. It also does not integrate with terrain awareness, so sometimes to avoid a collision with an aircraft it may suggest flying into a mountain (pilots are instructed to ignore TCAS RAs in this situation).

Memory bandwidth for unbatched decoder-only autoregressive inference should be the only bottleneck. The listed numbers imply you're only saturating about 500 GB/sec, ignoring KV cache traffic. Implies there's still quite a bit of inference performance left to squeeze out of the 960 GB/s spec.

One reason is that on-shore wind in the US is really good. The Great Plains are so windy and big that it's not really clear the added expense of marine turbines gets us much. Texas already has wind capacity factors of 40%, similar to many offshore projects.

Well that's gotta be an awkward family dinner.

Besides being power hungry and small, these fridges also had mechanical latches, which trapped and killed dozens of children per year.

This is going to shock people, but appliance manufacturers in the 1950s also had shareholders.

If GPT-4 is really 2 trillion parameters, that implies it uses at least 32 A100-class GPUs (that's with 8-bit weights). Assuming they're each consuming at half their max TDP, that's around 6 kilowatts. It seems to spit out around 15-20 tokens per second, so sampling 1000 tokens probably consumes around 100 watt-hours of power. That's an entire high-end laptop battery, or enough power to move a Tesla half a mile.

OpenAI probably have some magic to batch requests under high load but still, pretty power hungry. A good amount of the $0.06/1k tokens they charge is probably just to cover electricity.

Fp16 makes it easy to accidentally overflow, especially around summation operations.

It's interesting how much power we end up spending purely on DRAM transfer. HBM3 uses about 5 pJ/bit, so maxing out an H100's 3.35TB/s of bandwidth requires around 135 watts just for the RAM. Given how many AI workloads are entirely bandwidth-bound, you could probably build a much cheaper chip that just went all-in bandwidth and got rid of all the fancy compute elements.

So you're really happy that there is regulation requiring what the market already did on its own?

Looks like it would be $150/month for OP's needs?