On my way home, I noticed at a stoplight across the street from Apple Park that the driver in the lane next to me had his phone mounted up high in landscape mode and was watching the Simpsons. Just absolutely unhinged behavior lately.
Afaik, data center grade blackwell chips have never been legal for export to china. I think this has more do to with NVIDIA than DeepSeek. For a brief moment, people thought DeepSeek had found some way to produce AI without sending boatloads of cash to NVIDIA, causing a drop in share price.
Shortly thereafter people realized they were probably just evading sanctions and ~stealing~ bootstrapping parameters from other models to reach their stated training cost. This report is just further reporting on that rumor.
I'd put Noita in that category. I usually describe it as "broken both ways", because (as a rogue like) you have very little healing, and the enemies are punishingly hard. Not only that, but it's a full falling sand+physics simulation, so certain elements will randomly combine and kill you in the most unexpected and spectacular ways. On the flip side, the wand system is near turing-complete, and gets abused in the most crazy ways, to the point that you can do millions of damage per tick. One of the most chaotic and fun games I've played!
That makes games that aren't fun unless you're a wizard with the systems. They have their place, but I'm not a fan.
Personally, I think devs should embrace some stuff being broken. It's a single player game, it doesn't need balance. One of my favorite RPGs is FF8 precisely because you can trivialize the game if you engage with the character building systems. It feels awesome to stomp things with your broken party.
Perhaps "requirement" was a poor phrasing of the idea. "Feature?"
> Personally, I think devs should embrace some stuff being broken. It's a single player game, it doesn't need balance.
Exactly! Bravely Default & Octopath Traveler's job systems are built around the idea that the system should be breakable. BD2 even has a push-your-luck system that adds a multiplier for one-shotting multiple encounters in a row: you can get like a 50 percent rewards boost from random encounters if your team is able to "go infinite" against the current enemy mobs. And there are skills to remove damage caps, so you know they thought about it.
OT did kinda tone that down some to add the timing and break mechanics; if you land enough hits on enemy weaknesses, they lose their current turn, and go into a stun status next turn (but go first the turn after that, so no stun locking!). But you still get end game builds that max out boost points every turn; you just can't usually one shot bosses.
I think Disgaea fits into that class. The solution to every problem is "more levels", but the game is basically built around that. I think it's also fun to craft your own challenges out of the raw materials given to you... to get as many levels as quickly as possible, to win at minimum levels, to win with only X, to ignore Y and Z, etc.
I have O'Reilly subs available both from my employer and my local library. Doesn't fix the UI issues but does at least shift the ROI calculus.
There are some applications that try to export O'Reilly books into Kindle formats, but every time I've tried they've mangled a few tables, formulas or sidebars, etc. I should probably sell or hand down my kindle and find something more suitable to O'Reilly.
My go-to analysis for these sorts of places is net income per employee. Back in the day, IBM was hovering around $5,000. Today, Kyndryl is still around $5,000 (2025). But the parent company seems to be now at $22,000 (2024). For comparison: Meta is at $800,000, Apple is at $675,000, and Alphabet is at $525,000. And Wal-Mart, the nation's largest private employer, is around $9,250.
Now, probably part of that is just that those other companies hire contractors so their employment figure is lower than reality. But even if you cut the numbers in half, neither side of that spin off is looking amazing.
As I recall, it doesn't incorporate the historic price of stocks. So if I bought 1 share of Nvidia for $10 10 years ago, it'd say I had a net worth of $180 then, not $10 (as it uses today's stock price).
1. To the extent that creativity is randomness, LLM inference samples from the token distribution at each step. It's possible (but unlikely!) for an LLM to complete "pig with" with the token sequence "a dragon head" just by random chance. The temperature settings commonly exposed control how often the system takes the most likely candidate tokens.
2. A markov chain model will literally have a matrix entry for every possible combination of inputs. So a 2 degree chain will have n^2 weights, where N is the number of possible tokens. In that situation "pig with" can never be completed with a brand new sentence, because those have literal 0's in the probability. In contrast, transformers consider huge context windows, and start with random weights in huge neural network matrices. What people hope happens is that the NN begins to represent ideas, and connections between them. This gives them a shot at passing "out of distribution" tests, which is a cornerstone of modern AI evaluation.
> A markov chain model will literally have a matrix entry for every possible combination of inputs.
The less frequent prefixes are usually pruned away and there is a penalty score to add to go to the shorter prefix. In the end, all words are included into the model's prediction and typical n-gram SRILM model is able to generate "the pig with dragon head," also with small probability.
Even if you think about Markov Chain information as a tensor (not matrix), the computation of probabilities is not a single lookup, but a series of folds.
A markov chain model does not specify the implementation details of the function that takes a previous input (and only a previous input) and outputs a probability distribution. You could put all possible inputs into an llm (there's finitely many) and record the resulting output from each input in a table. "Temperature" is applied to the final output, not inside the function.
Re point 1: no, "temperature" is not an inherent property of LLM's.
The big cloud providers use the "temperature" setting because having the assistant repeat to you the exact same output sequence exposes the man behind the curtain and breaks suspension of disbelief.
But if you run the LLM yourself and you want the best quality output, then turning off "temperature" entirely makes sense. That's what I do.
(The downside is that the LLM can then, rarely, get stuck in infinite loops. Again, this isn't a big deal unless you really want to persist with the delusion that an LLM is a human-like assistant.)
I mostly agree with your intuition, but I’d phrase it a bit differently.
Temperature 0 does not inherently improve “quality”. It just means you always pick the highest probability token at each step, so if you run the same prompt n times you will essentially get the same answer every time. That is great for predictability and some tasks like strict data extraction or boilerplate code, but “highest probability” is not always “best” for every task.
If you use a higher temperature and sample multiple times, you get a set of diverse answers. You can then combine them, for example by taking the most common answer, cross checking details, or using one sample to critique another. This kind of self-ensemble can actually reduce hallucinations and boost accuracy for reasoning or open ended questions. In that sense, somewhat counterintuitively, always using temperature 0 can lead to lower quality results if you care about that ensemble style robustness.
One small technical nit: even with temperature 0, decoding on a GPU is not guaranteed to be bit identical every run. Large numbers of floating point ops in parallel can change the order of additions and multiplications, and floating point arithmetic is not associative. Different kernel schedules or thread interleavings can give tiny numeric differences that sometimes shift an argmax choice. To make it fully deterministic you often have to disable some GPU optimizations or run on CPU only, which has a performance cost.
Instead of a naive dense matrix, you can use some implementation that allows sparsity. If element does not exist, gets a non-zero value which can still be sampled. Which theoretically enables all outputs.
You're describing "temperature". That is usually done using the softmax function which cannot output zero for any element. In fact zero temperature is special cased, or they do exactly what you just said (add a teeny tiny epsilon to everything) in order to avoid having to treat zero temperature as a special case.
Yes it is, there's no rule saying you can't manually mess with the state transition function while designing and building the chain. It's still a fixed function at runtime. It's just that you don't need to do so in the first place, because this is done already after the markov chain has been evaluated, if temperature is set to nonzero (and some systems don't even really allow you to use zero temperature, they only allow you to make it really small)
One of the reasons why I changed banks. My new bank has a coin counting machine in the lobby, you throw your coins in, it consumes them, and gives you a slip that you take to the teller.
As I understand it, coins are considered a government service. Banks and retailers pay to deal with them. Buying them from the public for face value actually saves them money.
It's so easy to use coins, pennies included, in day-to-day transactions I never accumulate any. Accumulating pennies or other coins is a concept I don't understand. You can spend up to 4 pennies in any purchase you do, and if you don't can't never receive more than 4. For nickels, dimes, and quarters, the maximum is smaller.
If a person has good basic arithmetic skills and it is a priority for them, then yes they can use coins easily. However, a lot of people either can't do the math or are unwilling to use change correctly.
For myself, it's such a priority that I'll get upset with myself if I have more than 4 pennies.
Japan has more coins (in regular use) than USA, so giving the correct amount is even more important or you'll end up carrying a lot of coins. 1000 yen is the smallest bill so... Example: 999 yen. 500,4x100,50,4x10,5,4x1 yen coins, 19 coins total.
When I used to use cash I used to do this all the time. I would nearly always overpay to minimise the number of coins in my pocket. For example I had a bill of £1.63 and I was paying with two £1 coins, I would get 37 pence in change, which would be a minimum of four coins (20p, 10p, 5p, 2p). So I would pay £2.13 to get a 50p or £2.03 to get two 20p coins.
95% of the time the person serving me would clock on to what I was doing but the other 6% it'd take some persuasion, and occasionally they would insist on giving me back my overpayment before ringing it up.
Most of our supermarkets have at least one self service machine that accepts change. Once a week I pour any loose change in then settle the rest with a card.
When I lived in Scotland there was a "loose change" machine at the local Tesco. You pour in your coins and it would give you a receipt you could take to a cashier to get cash back - but the downside was that it charged you something like 10% of the total as a fee. Which I wouldn't pay.
Edit: I just searched and the Tesco documentation says "There is a 25p transaction fee and an 11.5% processing fee on the total amount of coins you put in the Coinstar centre. For charity donations, this processing fee is reduced to 8.9%." (wow, how generous!)
Back in the day, I'd sift through my jar of change and keep the quarters, which were good for parking meters and laundry. The rest went into the Coinstar machine. The fee for counting dimes, nickles, and pennies seemed OK.
The machine always had some weird foreign coins or subway tokens left over by the previous customer in the reject bin, which was potentially interesting.
*You don't think I'd go into combat with loose change in my pocket, do you?"
But I must admit that I never formed the habit of bringing change with me when I go somewhere. So it piled up at home. The quarters were easy: They got saved for parking, laundry, etc. But I ended up with a sack of pennies that I finally cashed in at the bank.
It's odd how banks have largely stopped operating change counting machines.
In my childhood we'd hoard loose change then make a trip to the local po-dunk bank serving my neighborhood surrounded by corn fields, and even there they'd take our bucket of loose change and dump it into a counting machine for free.
It was a game to try guess the amount we'd get in paper cash...
Now you have to pay for this service at a grocery store using a cumbersome machine operated by Coinstar.
COVID happened. However, all three of the banks I visit regularly (over branch of a national bank, two branches of a local credit union) all have coin counting machines in the lobby, though it took awhile for them to be added back to the branches that took theirs out.
No doubt COVID kicked skimpflation into high gear, but this was already a pattern I noticed long before 2019.
It seemed to generally coincide with the demise of retail in general, and of course the elimination of bank-teller interactions and emergence of ATM machines. All of these things are a blurry mess from my past...
Average gift card has a discount of 8 to 20% built in. Looks like Coinstar is currently charging 12.9%, so a gift card could actually be more profitable for them.
Mine had the machines, then ripped them out, over the cost to them the regional bank they deal with imposed and other excuses. Coinstar (some) gift card is the only no-fee I've found in my area, but then you're stuck with a gift card instead of cash.
I have talked to my bank and was told not to roll them, they just throw them in a machine to count them and deposit the money in my count. It is not uncommon to see people bring in a box of coins and the bank takes care of them.
I have like 2 dollars in coins, not even a roll of pennies. I just thought I'd try depositing it while running a different bank errand and they were like "naw, go to coinstar with your poor person money"
In Canada, I've only ever seen these in grocery stores, operating for a fee (and they don't accept commissions) and a singular credit union branch (because they serve the underbanked at that particular location).
It seems more reasonable than the outright refusal of many businesses to accept cash at all, and plus this transaction isn't even a "debt" to which the penny would be legal tender.
As I understand it, more than X dollars worth of coins is not legal tender. I learned this due to an absurd case in Detroit, where someone stole bags of coins from an armored car, got caught, and claimed their crime was not a felony because it was below the dollar limit for a felony. Of course the judge treated their request with the disdain that it deserved.
There are multiple laws that could have been broken to make it a felony, but if the only reason it would have been a felony was the dollar amount, I'm actually less inclined to side with the judge.
This is all third-hand, through a game of internet-telephone, and my money (if you'll pardon the pun) is on there being additional factors though
reply