JonathanFly

Suno is kind of underselling this with their default voices. With just a little effort you get great stuff, very emotive, fantastic cadences. Here's my female David Attenborough I was playing around yesterday. Not the clearest but charming.

https://user-images.githubusercontent.com/163408/238257231-4...

And Bark is more than TTS. While I haven't had much success with one-shotting full songs with Bark, you build a decent sample library with it. A few recent ones I saved, just randomly concatted here:

https://soundcloud.com/jonathan-fly-620508219/bark-all-night...

>Am I hallucinating or didn't several of the examples have background audio artifacts, like it's been trained on speech with noisy backgrounds, I'm guessing audio from movies paired with subtitles? Having random background audio can make it quite hard to use in production.

The other side of that problem is an opportunity. That's why the same model can also generate music, background noise and sound effects. And it's just because the prompt specifies those things explicitly. The input is truly semantic, so the output is rich and reflects that context. Is your input text sounds like it came from a speech, then there's a high chance your output audio will sound like a megaphone in a public space with crowd reactions and maybe even applause.

Totally agree. The samples in the OpenAI paper didn't seem that impressive and there were so few of them, I assumed it must not be very good. But even the one-shot version of the scenery you linked is great. Very efficient use of SVG features.

Look at this majestic beast GPT-4 just drew! https://imgur.com/a/ogZ6qI0

Interestingly, I think you often get better results with ChatGPT web interface than the API (the OP used a system message), but for a funny reason.

Because ChatGPT typically first responds by insisting it can't actually draw, it ends up accidentally chain-of-thought prompting itself into drawing better. "I can't draw, but I can describe a detailed drawing for you in words." So instead of winging it with the SVG code and getting confused, it goes into it with a plan.

This is how it described the unicorn I linked:

__User__ Draw a unicorn!

__ChatGPT__ As an AI language model, I am unable to create visual images directly. However, I can describe a detailed image of a unicorn for you in words.

Picture a magnificent unicorn, standing tall and proud in a serene setting. The unicorn's coat is a shimmering, pure white, and it glistens under the warm sunlight. Its body is strong yet elegant, with a graceful arched neck and powerful, slender legs. The hooves are cloven and emit a faint, otherworldly glow.

The most striking feature of the unicorn is its majestic spiral horn, which emerges from the center of its forehead. The horn is long, slender, and gently tapers to a sharp point. It is iridescent, with colors swirling together like a mesmerizing opal. The unicorn's mane and tail flow behind it like silken strands, shimmering with an ethereal pastel rainbow of colors.

Its eyes are large, expressive, and full of wisdom, with a depth that speaks of untold secrets and ancient magic. The unicorn's ears are pointed and alert, picking up the slightest rustle of leaves or the gentlest of whispers. The unicorn's expression is one of serenity, exuding an air of peace and tranquility that surrounds it like a gentle embrace.

Though it is a creature of great power, the unicorn carries itself with a sense of humility and grace that inspires awe and respect in all who behold it.

__User__

What's the SVG for the image you just described?

__User__

continue (to finish the file)

Some more examples here, but with many iterations of "Improve this" over and over:

https://twitter.com/jonathanfly/status/1639657098605740035

Wow, I can't thank you enough for this. Somehow I never noticed that GPT-4 doesn't use separate tokens for each tab like 3.5. I was wasting a lot of time minimizing excessively tabbed code to save on the token count! Like seriously way too much time, all based on a bad assumption.

https://twitter.com/jonathanfly/status/1643633463260577794

Super cool. I wish he'd posted more of the prompts, part of what makes so impressive is how simple they are, like the chess example.

Old hardware device with non upgradable memory and only a SATA interface? Like old game consoles.

I have something myself that is in the ballpark. It's e-ink Thinkpad that has 4GB of soldered memory. It would be plenty fast if only it had more memory. They don't make a newer e-ink laptop like this. It's not SATA but if there was a way to stick a 32GB DDR4 memory stick onto an NVME drive, use it in the second slot as a pure swapfile, that would be a cool little trick to extend the lifespan of this thing.

I might try this with Optane if the prices come down more.

I love this. I first tried it with Grover in 2019 and ran into similar token issues: https://twitter.com/jonathanfly/status/1143773270451396608?l...

Looks like you don't have a 'most common/simple words' vocab restriction? That's a pretty fun one too. https://twitter.com/jonathanfly/status/1144345934857474048

The capability of a language model I care about most is probably its ability to represent or simulate Captain Picard. In the sense of being good at creative tasks but also Captain Picard, specifically. Is OpenAI deliberately doing something different on purpose that makes their models better for this, or is just that OpenAI has a lot more copyrighted data in their dataset, as I noticed just now when skimming the Facebook paper for MMLU section and seems be what the Facebook folks think?

"A potential explanation is that we have used a limited amount of books and academic papers in our pre-training data, i.e., ArXiv, Gutenberg and Books3, that sums up to only 177GB, while these models were trained on up to 2TB of books. This large quantity of books used by Gopher, Chinchilla and PaLM may also explain why Gopher outperforms GPT-3 on this benchmark, while it is comparable on other benchmarks."

While it's technically not a bad price in terms of raw compute and memory, the noise from the high speed server fans alone is reason enough not to get this. Never mind the power usage.

I was going to say maybe it'd be worth it if they were 24GB GPUs, but I'm not sure you can even use recent Pytorch with cards that old. You'd have to work around the limitations.

You don't even need a GPU to learn anyway, you can use tiny models that train super fast even on CPU for that. You need the beefy GPUs once you want to generate tons of content or train a modern model on big datasets.

Get a 3060 12GB or a 2080Ti 11GB and call it day, at most.

>A VSCode Extension to make your code worse with AI

>Document the code with emoji

>Document the code in any style (Dirty Limerick, Fast Talkin' 1930s Gangster)

>worse

TIL I've been coding wrong my whole life. But it's the only way I know...

>I still quite don't get it. What are hustle bros exactly? Just people who chase short-term money?

The YouTube thumbnails in the article might help. Before ChatGPT these videos would be about drop shipping and NFTs.

https://duet-cdn.vox-cdn.com/thumbor/0x0:3840x2160/1200x675/...

Magician Smart Status reports 0 'media errors' on my 980 2TB. Is that suspiciously low or does it mean things are actually ok? Is there a different tool I should check with?

Not bad with a single out of context chapter from a genre web fiction serial that is hard to understand without context. The language and writing style sounds a LOT like ChatGPT. Feels like the model asked a bunch of different questions (what is interesting to note, what's important, etc.)

>Jessica Yamada is a therapist working with a patient, Victoria, who is struggling with the possibility of uncontrollable, incurable plagues. Jessica is also dealing with a malfunctioning suit that has interrupted her sessions with her patient. She is also dealing with a colleague who is making her out to be a freak. Jessica is also trying to help a young woman, Lily, who is torn between fatalism and wanting to help her friends stick together. Finally, Jessica is talking to Eidolon, a powerful man who is losing his powers and is trying to fight a monster to the death. It is interesting to note that Jessica is trying to help her patients in a variety of ways, from providing coping mechanisms to offering moral support.

https://parahumans.wordpress.com/category/stories-arcs-11/ar...

Also, unintentionally got this line perfect (given later story stuff...)

'a powerful man who is losing his powers and is trying to fight a monster to the death'

According to some people (Justin Sung on YouTube is where I heard it, probably, so no idea if credible) you can influence the forgetting-curve and make Anki more effective by having more context, putting knowledge into relationships with other knowledge, etc. It's a multiplier on raw/rote space repetition.

Totally agree. But I'm tempted solve this by adding MORE AI. Have the tool prompt GPT to quiz you about the new card, or force you to use a new language vocabulary word a bunch of times. Maybe even argue against you adding the card and try to convince it it's worthwhile information.

USER: Add a new Anki card for, "It is important to set a high standard from day one."

GPT: That seems like generic advice. Of course high standards are better than low standards in a vacuum, but everything has a cost. Being too rigid in setting high standards may limit the business's flexibility to adapt to changing market conditions, customer needs, and technological advancements. Are you sure high standards are that important?

(That criticism actually generated with GPT, using the example the author of the tool Tweeted.)

But most projects aren't designed to use it that way, as a virtual 48GB single GPU.

All the frameworks work on CPU. At the time I tried it, the 5950X was about 10x slower than my GPU, which was a 1080Ti or 2080Ti. GAN not a transformer though.

Basically 'it depends' is the answer to all your questions, but dual 3090s is a perfectly fine choice. Though ideally you would have NVLINK since it is an advantage over the 4090. In some specialized situations it is possible to have NVLINK act a lot like 48GB of memory, but if you don't already know if you can leverage NVLINK, you very likely aren't in that situation.

It's an interesting option. My gut instinct is that if you need need 128GB of memory for a giant model, but you don't need much compute - like fine tuning a very large model maybe - you might as well just use a consumer high core CPU and wait 10x as long.

5950X CPU ($500) with 128GB of memory ($400).

Any chance you will support VR? Having tried a flat screen virtual pinball table, the VR experience just blows it away.

Or PS5 DualSense force feedfack for haptics and feeling the ball?

>It's amazing to me how virtual and real pinball are really not comparable at all in terms of enjoyment (at least for me personally). I still don't really understand why.

Have you tried it in VR? With free head movement, and proper 3D sound simulation, the physics feel a lot more real.

It's still missing good haptics. I think the PS5 DualSense force-feedback controllers and trigger might be able to give you the feeling of the ball rolling around different places, but I don't see anyone trying to make this work in virtual pinball yet.

Silly random thoughts that went through my head at this question

1) Many people's life savings is 0. A bargain!

2) If _I_ were one of the five people, I'd try and make it up to the lever puller. Lever-guy might even come out ahead.

3) If I pull this lever right away in question 3, if there are later questions about giving up my life savings, they'll be free. This might be the value lever pull! I might be kicking myself later for missing this deal.

If it were truly the one single opportunity I will ever have to choose to stop existing, it's a tougher choice for sure.

But on the other hand, I might be so curious about what in the heck happened in the world outside - the world that led to this universe where it will be impossible for me to die for eternity, and somehow I know this with absolute 100% certainty - that I still opt out.

>Oh no! Due to a construction error, a trolley is stuck in an eternal loop. If you pull the lever the trolley will explode, and if you don't the trolley and it's passengers will go in circles for eternity. What do you do?

50% of people pull the lever?!?

I hereby declare that if I'm ever going to be stuck in a trolley for my entire life, I do NOT want the lever pulled. Toss me a smartphone charger and my life wouldn't even be that different, day to day.

>What to do with all that computing power now. I guess AI.

For AI, the lack of memory is pretty limiting.

Having a ton of 8 GB or 10 GB memory GPUs (3070/3080) is annoying and often just not even possible to use effectively without designing specifically to it. If they get really really cheap maybe they'll be some tools or models that are built to leverage scenarios like that, it's not impossible, but for now it's probably just the cheap 3090s (24GB of memory) that will benefit AI.

>I gave up on plants versus zombies 2 quickly when it became obvious the game was just prompting me to spend money at every turn. Granted, some people say this was mitigated eventually.

The funny thing is the unmitigated, most pay2win version of Plants Versus Zombies 2 - the game at launch - was accidentally one of the most amazing and intense gaming experiences for people who never spent anything. For people who just really enjoyed a vastly more challenging version of PvZ.

I loved the first game but it was never a challenge, it was rare to fail a level more than once. But the levels in the second were passable if you spent zero premium currency (even not spending any of premium currency they gave you for free for logging in) - they were just very very challenging. Passing every single one required novel strategies and slowly revising, eking out a few more tiny edges each time you try, combined with absolute precision in execution, until somehow the level was passable.

I'm not actually sure every single level was passable by not even spending the currency they gave you for free, I never quite finished it, but I got through most of them! When I logged in to finish the remaining levels later, they had made them all way easier in an effort to respond to p2w criticism so I never had a chance.

The writer of Severance specifically mentioned the Stanley Parable, so that's quite intentional: https://www.inverse.com/entertainment/severance-apple-tv-plu...

The article is about how these coups are reducing US influence and strategy.

The sub headline: "Insurrections are disrupting American security strategy in the region and giving Russia an opening to gain sway."

Well, this beats the version of Dune on my HP48GX that blew my mind in the 90s.

I wish there was a good dedicated portable Pico-8 device. The POCKET C.H.I.P. was a great idea but the keyboard and buttons were way too uncomfortable to play for more than a few minutes.