So to follow up on this, I've caught it in action. Models, when quantized a bit, just do a bit more poorly with short contexts. Even going from f32 (as trained) to bf16 (as usually run) to q8 tends to do okay for "normal" context windows. And q4 you start feeling like "this model is a little stupid and gets stuck sometimes” (it is! It's just that it's still mostly careening about in the space of "plausible" most of the time. Not good guesswork, but still in the zone). With long contexts, the probability of parameters collapsing to zero are higher, so the more context the more likelihood you are to see brokenness.
And then at Q2 (2 bits per parameter) or Q1, the model falls apart completely. Parameters collapse to zero easily. You start seeing "all work and no play makes jack a dull boy” sorts of behavior, with intense and unscrutinized repetition, followed by a hard stop when it just stops working.
And quantization is a parameter that a model vendor can turn relatively easily. (they have to regenerate the model from the base with more quantization, but it's a data transformation on the order of running a terabyte through a straightforward and fast process, not like training).
If you have 1000 customers and enough equipment to handle the requests of 700, going from bf16 to q8 is a no-brainer. Suddenly you can handle the load and have a little spare capacity. They get worse results, probably pay the same per token (or they're on a subscription that hides the cost anyway so you are even freer to make trade-offs. There's a reason that subscription products are kinda poorly described.)
It's also possible for them to vary this across a day: use models during quieter periods? Maybe you get an instance running a bf16 quantization. If you use it during a high use period? You get a Q4 model.
Or intelligent routing is possible. No idea if anyone is doing this, but if they monitor what you send a bit, and you generally shoot for an expensive model for simple requests? They could totally substitute a highly quantized version of the model to answer the question.
There are •so many tricks• that can be pulled here. Some of them very reasonable to make, some of them treading into outright misleading or fraudulent, and it's weirdly hard to draw the line between them.
Very proud and excited to vote in the NDP leadership race today!!
This is not the first time I've voted in a Federal leadership race... more on that later but first, my choices! I considered only voting for two people, but I ended up filling in all 5 choices.
#1: Tanille Johnston @…
#2: Avi Lewis @…
#3: Heather McPherson
#4: Tony McQuail
#5: Rob Ashton
Why?
You might ask why I would publicize my choices. I don’t expect others to of course. It is a privilege and a right in Canada to exercise your democratic choice freely and privately, but I also think there is value in knowing how others voted.
#1 why Tanille? #electoralReform and proportional representation myself, I didn't just want to pick my top two. I wanted to make a statement on each of these candidates an influence each one.
To be blunt, Heather is #3 because she is the middle-of-the-road candidate. She is an excellent representative as MP and has gathered the support of other MPs including my own, but while I would be OK with her leadership, I would see her as a continuation of the status quo, and that is not what the NDP needs as a party, nor is it what Canada needs as a country.
We desperately need a vigorous and clear alternative to the Centre-but-mostly-Right Liberals, and the MAGA-wannabe Conservatives. The only way to do that is to catch the attention of Canadians and inspire them. I am not sure that Heather has the ability to do that, and if we continue with the same leadership crew in the NDP, I am not confident that the policy choices will be strong enough to inspire and attract Canadians.
That is why Tanille and Avi are far better options.
#4 Why Tony:
Tony is the real deal. Honestly, I would have loved to rank him higher. He represents the true life blood of rural, socially progressive, environmentally aware, Canadians. You should go check out his platform. I am so glad that he was able to participate fully in the race and we need his voice in the NDP.
#5 Why not Rob?
I have been an active member in my Union for more than 10 years. Unionism is The Way. Rob is representing a division within the union movement that claims that working people can't have jobs if the environment is put first. This is a lie.
We need union leaders that look to the future and speak honestly to people. We need union leaders who are genuinely progressive, not ready to do the bidding of corporate masters to the benefit of a few.
Working people need honesty, and when an industry is on decline, a clear path to new, excellent, union, jobs!
#CanPoli #CdnPoli #Liberal #CPC #Canada #Democracy #NDP
Fernando Mendoza Gets Major News Amid Raiders Draft Link https://heavy.com/sports/nfl/las-vegas-raiders/indiana-fernando-mendoza-major-news/
Daleks, in the future, are teaming up with the heads of the other galaxies to overtake the Solar system and destruct time, and the Doctor's only got Steven (a pilot from the 24th Century) , Katerina (a slave girl from ancient Troy), and a local soldier to help.
The guardian of our Solar system has betrayed us to the Daleks! He's mined 50 years worth of Terrainium secretly from Uranus to power the core of the Dalek Time Destructor.
The Daleks say "Execute" when they have found someone guilty of negligence, vs just when they are a pest to be exterminated.
The doctor nips in, under disguise, to investigate the council, steals the Terranium and the president's ship, then gets the team stranded on the Solar system's prison planet.
The prisoners try and raid the ship but the Doctor has set a trap and electrocutes the invaders, just in time for them to fix the ship and escape.
Only one prisoner has stowed away on board.
[Then there's a episode still missing, in which apparently Katerina wrestles the prisoner into the air-lock and they are both spaced. The Doctor and Peter return to Earth to warn about the Daleks.]
They arrive on Earth (future earth remember, but all the computers have giant tape drives and knobs) as an experiment on mice is in progress.
I guess the experiment was to try and make mice turn into negative images screaming in slow-motion and then bounce up and down as they are transmitted through space many light years away. And the Doctor, Steven, and some security guard chasing them get sent along too. With the Daleks following on in their ships.
The Daleks exterminate the mice 😔
There's 8 ft tall invisible creatures on this planet so the mice were gonna be in trouble anyway. The Doctor beats them off with sticks before being apprehended by Daleks.
[Then there's four still-missing episodes in which the Doctor and Steven steal a Dalek ship, trick the Daleks with a fake Terrainium core, meet the Monk who attempts revenge, and celebrate Xmas on a silent film set. All with Daleks giving chase]
The security guard and the Monk are still with them in the next archived episode, when they are in a Egyptian tomb for some reason and the companions including the monk are captured.
The doctor faces the Daleks to negotiate his companions' return.
At the hostage exchange the Doctor hands over the core as the ancient Egyptians attack the Daleks. It's a slaughter of course. All the Egyptians die, but they made a good distraction and the Doctor skips off.
He's knicked the Monk's Tardis' directional compass so the Monk goes to who knows what random place now.
The Doctor aims to try and materialize the Tardis at the point the Daleks are likely to use that Terranium, to take over the galaxy and destruct time, but seems like the Tarids fails.
[And then there's another two still-missing ones in which the security guard ages to death in a time-mishap, and an entire planet is wiped of all life to thwart the Daleks. The Doctor and Steven lament the senseless deaths of the three of them that they cared about.]
Crikey. I guess they used to bounce around in time and space more during a story when it was twelve 20 minute episodes. That Prison Planet was there only to be landed upon, have the Doctor electrocute some people, and then leave with a stowaway. The 8ft tall invisible creatures are in like 2 scenes.
Incredible body counts. Just absolute carnage compared to most New Who.
The background of mega-death while the protagonists lament the death of only their own reminds me of the way the contemporary news will focus on one marooned soldier over the deaths of hundreds. Humanize only their own.
The Monk is a good candidate for a return. He's got this great Frankie Howerd like mischievous campness. Exited this story with a randomizer on his tardis vowing revenge.
#watching #tv #doctorWho #TheDaleksMasterPlan