1,431 karma · joined March 14, 2016
The same way everyone else fails at it.
Change some hyper parameters to match the new hardware (more params), maybe implement the latest improvements in papers after it was validated in a smaller model run. Start training the big boy, loss looks good, 2 months and millions of dollars later loss plateaus, do the whole SFT/RL shebang, run benchmarks.
It's not much better than the previous model, very tiny improvements, oops.
Pre-training: You train on a vast amount of data, as varied and high quality as possible, this will determine the distribution the model can operate with, so LLMs are usually trained on a curated dataset of the whole internet, the output of the pre-training is usually called the base model.
Post-training: You narrow down the task by training on the specific model needs you want. You can do this through several ways:
- Supervised Finetuning (SFT): Training on a strict high quality dataset of the task you want. For example if you wanted a summarization model, you'd finetune the model on high quality text->summary pairs and the model would be able to summarize much better than the base model.
- Reinforcement Learning (RL): You train a separate model that ranks outputs, then use it to rate the output of the model, then use that data to train the model.
- Direct Preference Optimizaton (DPO): You have pairs of good/bad generations and use them to align the model towards/away the kinds of responses you want.
Post-training is what makes the models able to be easily used, the most common is instruction tuning that teaches to model to talk in turns, but post-training can be used for anything. E.g. if you want a translation model that always translates a certain way, or a model that knows how to use tools, etc. you'd achieve all that through post-training. Post-training is where most of the secret sauce in current models is nowadays.
If a mod includes malware it'll be very obvious as mods are usually small.
PS5 + 3 years of PS Plus = $740
Steam Machine = $700
Add/remove more years of PS Plus if the SM turns out to be more/less expensive.
If you add the fact that games on PC are usually cheaper and have sales more often then it's a no brainer, but that won't convince the FIFA and COD players.
This can be done with finetuning/RL using an existing pre-formatted dataset, or format based RL where the model is rewarded for both answering correct and using the right format.
They come preloaded with a pirate iptv service that only works for 1-2 months then they ask you to pay something like $70/year to keep watching. There's tons of providers for these IPTV services so bundling them with the boxes is a way to make it easy to access while gaining subscriptions, you can just buy a cheap android TV box yourself, install the apk and get a cheaper IPTV provider.
Most of these boxes/providers don't last more than a couple years as authorities tend to go after them when they get too big. My dad uses them to watch portuguese TV--it would be impossible to watch certain channels outside the country otherwise--and in the past 10 years he changed provider 3-4 times.
IIRC the wisdom of the time cloud started becoming popular was to always be on-prem and use cloud to scale up when demand spiked. But over time temporarily scaling up became permanent, and devs became reliant on instantly spawning new machines for things other than spikes in demand and now everyone defaults to cloud and treats it as the baseline. In the process we lost the grounding needed to assess the real cost of things and predictably the cost difference between cloud and on-prem has only widened.
I stopped watching him completely around the time of the intel dGPU release. He would show leaked roadmaps of intel's dGPU launch with Celestial and Druid on there, but the video would be him basically repeating the narrative that the division is on the verge of cancellation and has no future, etc. The documents he has leaked almost never match the titles and narratives he pushes. He's not always wrong, but his biases are clear and the titles are more often misleading clickbait than factual.
Tom Petersen (Lead Intel GPU Engineer) has showed up on multiple interviews with LTT and GN among other tech channels and has talked at length the things his team did on the current gen, with heavy implications of what's coming up next (as far as he can without breaking NDA). His in depth analysis of GPU architecture are a much useful use of my time than listening to a guy that was given 2 leaked slides of a 6-month old powerpoint speculate how it spells doom for whatever company.
If you think logically, it makes zero sense to cancel Celestial now. According to Petersen Arc's hardware team has been working on Druid for months now, and unless the software team is severely behind with the drivers and support then at the very least Celestial will receive a limited release. They already did a limited release of Battlemage to put more resources on Celestial, it would be a shame to throw all that effort away now.
Alchemist - First gen GPUs A310 GPUs are the low end, A770 are the high end. Powerful hardware for cheap, very spotty software at release. Got fixed up later.
Battlemage - Second gen (current gen), only B570 and B580 GPUs came out. They said weren't gonna release more Battlemage GPUs after these because they wanted to focus on Celestial, but probably went back on it seeing how well the B580 was reviewed and the B770 is due to be released by the end of the year.
Celestial - Next gen GPUs, they were expected for release early 2026. This article claims it was cancelled, but personally I think it's too late to cancel a GPU this late in production. Especially when they basically skipped a generation to get it out faster.
They're a channel focused on leaks, but most of their leaks are just industry insider gossip masked as factual to farm clicks. Their leaks are useless for any sort of predictions, but may be interesting if you'd like to know what insiders are thinking.
A quick google search also yielded this[1] 2-year old reddit thread that shows videos they deleted because their predictions were incorrect. There's probably many more. (That subreddit seems to be dedicated to trashing MLID.)
[1] https://www.reddit.com/r/BustedSilicon/comments/yo9l2i/colle...
One was the first gen Alienware M17 with two GTX 270M GPUs (yes two) and an onboard nvidia GPU whose specific model I can't remember. That one suffered from stutters and audio crackling, etc. It was sort of fixed by disabling SLI and the onboard GPU and sticking to a specific driver that was modded, the driver was by someone on the notebookcheck forums IIRC. Later on I think it got somewhat patched with a bios update that let you use SLI without the stutters, but I think the laptop reached EOL without it being fully fixed.
The second was an ROG ASUS laptop with a GTX 460m (I can't recall the laptop model). Pretty much the same story as the OP but I didn't have the knowhow to go deep into the ACPI code. The only change from the story is that latencymon kept attributing the latency spikes to multiple dlls, sometimes it was some wifi driver, other times it was an nvidia one. I don't remember the full fix for that one, but it involved me changing the wifi card and disabling the dGPU (not the onboard one) when I was not gaming so I could watch videos and such without it crackling. Funnily enough it didn't crackle much when actually playing games (it still happened, just very rarely).
I stopped buying gaming laptops after that. Seeing this story makes me think things haven't changed one bit.
I put docker-compose files in ~/configdata/_docker/
The docker-compose files always mount volumes inside the ~/configdata/ directory. So let's say irssi has a config directory to mount I'd mount it to ~/configdata/irssi/config/
Then I can just run a daily backup on ~/configdata/ using duplicati or whatever differential backup tool of your choice and be able to restore application state, easily move it to another server, etc.
Hopefully this means that a third player will join and provide a truly open android platform.
If an OS is owned and controlled by a single company, it's never truly open. No matter how much they claim it is.
edit: I should clarify, I'm talking about devs that develop for third party stores exclusively. Usually privacy conscious or devs whose apps aren't allowed on the play store for this and that reason like tachiyomi.
Also note that these are not barriers to being understood, but they are barriers to be fluent in the language. These people have lived in the US for 10+ years and communicate in english just fine, humans need very little language to communicate most things. But if they need to be taken seriously in jobs that require constant communication, becoming fluent should be a long term goal, and outputting early is bad. It's best to wait for 1-2 months until you get a grasp of sound and flow of the language.
Writing and speaking are effective at establishing long term memories, it's why we do it for other things, but a language learning beginner has no idea if what they're writing makes sense or if there's any subtle mistakes in how they're pronouncing words or how they're putting them together, etc.
Language learning experts don't recommend you start speaking/writing unless you have a coach or have reached an intermediate level so that you can discern when something sounds native or not. That way you can self evaluate with recordings, etc.
It is an effective tool for learning, but for self-learning you're gonna be shooting yourself in the foot long term. You should only do it if you have, say, a partner that speaks the language and doesn't mind correcting you all the time.
For Japanese I recommend that you do learn how to write kana/kanji from the start, and even some vocab if you want. But stop there. Don't write sentences, don't try to talk to japanese people on those apps/discord etc. and wait until you're at an intermediate level to do it, otherwise you'll form some very bad habits that are very hard to undo.