thorum

Someone probably said the same thing to Van Gogh during his life.

OpenVoice currently ranks second-to-last in the Huggingface TTS arena leaderboard, well below alternatives like styletts2 and xtts2:

https://huggingface.co/spaces/TTS-AGI/TTS-Arena

(Click the leaderboard tab at the top to see rankings)

Children of Time (Adrian Tchaikovsky), The Expanse (James S. A. Corey), Broken Earth (N. K. Jemisin), Locked Tomb (Tamsyn Muir), A Memory Called Empire (Arkady Martine), Murderbot (Martha Wells), The Light Brigade (Kameron Hurley), The Wayfarers series (Becky Chambers), and The Calculating Stars (Mary Robinette Kowal) are some of my favorites. Brandon Sanderson’s Skyward series is also worth a read.

The article’s premise is based on a linked Washington Post survey that shows 12% of readers read scifi in the last year, fantasy 15%, and romance 11%. This is not compelling evidence that scifi is unpopular. Perhaps people are just more diversified in their reading habits now across the board. It would be nice to see some more in-depth data showing trends across time before we leap to the conclusions this article does.

> the researchers collected a sample of 1,920 video clips from Twitch.tv. These clips were selected based on their popularity within the most frequented categories of the platform. The sampling process was carried out from September to October 2022, focusing on the top videos within the videogame and “In Real Life” (IRL) categories, which were further broken down into subcategories like “JustChatting,” “ASMR,” and “Pools, Hot Tubs & Beaches.”

This isn’t measuring whether female streamers are sexualizing themselves. It’s measuring what kind of clips get the most views from Twitch’s audience of mostly young male viewers. It shows that they watch a lot of clips from women using Twitch to advertise their OnlyFans pages, who tend to stream in the categories listed.

There are many popular female Twitch streamers who just play video games (in categories dedicated to video games) and are not sexualizing themselves at all.

Note that the benchmarks used for comparison are basically measuring the model’s ability to understand financial content. In other words, reading comprehension for English, just in a specific domain. It shouldn’t really be surprising that a strong generalist model performs well here.

On the other hand, GPT-4 actually did worse on the NER task - labelling and tagging terms used in the text - vs their finetuned model. I assume the finetuned model was better at using the specific labels they were targeting.

I don’t know what the people who say Claude 3 is better than GPT-4 are using it for. It’s been consistently worse for everything I’ve thrown at it.

Debugging a Python function this morning. Claude 3 Opus failed completely. GPT-4 found the bug, as well as two others I hadn’t even been looking for.

That might be true for finetuning ChatGPT 3.5, but if you can finetune a small model (7B or less) to perform on par with GPT-4, while being faster and private, that’s a different story.

> The research team demonstrated the system’s accuracy by having the participants pronounce five sentences — both aloud and voicelessly — including “Hi, Rachel, how are you doing today?” and “I love you!” (…) Going forward, the research team plans to continue enlarging the vocabulary of the device through machine learning and to test it in people with speech disorders.

It's a proof of concept at this stage but very cool.

My first thought on seeing this was, I can’t believe the Belters never thought of this in The Expanse.

The 2012 surgeon general report linked in the article goes into depth about the debate on this topic and the research that has been done into the effects of smoking in films on children:

https://www.ncbi.nlm.nih.gov/books/NBK99238/#ch5.s78

I’d be fine living in a world where nobody bothered to invent non-consensual mind reading tech. Or, if that’s too much to ask, where we invent everything else that needs inventing first.

From the article,

> The present work demonstrates that it is now practical for patients to undergo a single MRI scanning session and produce enough data to perform high-quality reconstructions of their visual perception.

> Such image reconstructions from brain activity are expected to be systematically distorted due to factors including mental state, neurological conditions, etc.

> This could potentially enable novel clinical diagnosis and assessment approaches, including applications for improved locked-in (pseudocoma) patient communication (Monti et al., 2010) and brain-computer interfaces if adapted to real-time analysis (Wallace et al., 2022) or non-fMRI neu-roimaging modalities.

> As technology continues to improve, we note it is important that brain data be carefully protected and companies collecting such data be transparent with their use.

It’s impressive how well the T5 family of models has aged, even compared to newer LLM architectures.

The submitted link works for me, but maybe it’s a regional issue. Here’s an archive.org link:

https://web.archive.org/web/20230329042339/https://www.colum...

Is it not possible (yet) to return them to the surface in a controlled manner via a SpaceX capsule of some kind? (Or just not seen as worth the expense?)

I’m not a quantum computing expert, but surely it’s more likely the algorithm is less original than the authors’ believed, than it is that Opus is doing original scientific research.

If this were correct, it would be a bit less impressive than OP’s claim, but still a monumental leap forward for translation of low-resource languages - a task which all previous LLMs fail at.

Full benchmarks vs other Mistral models:

https://docs.mistral.ai/platform/endpoints/#benchmarks-resul...

Supported file types: https://github.com/google/magika/blob/main/docs/supported-co...

xtts2 is great, but it looks like this model is probably more consistent with its output and has a better grasp of meaning in long texts.

So how many of these would you need to match the brightness of a lightbulb?

I asked it to do a natural language processing task and it said it couldn’t because it was only a language model.

I mostly agree, except if you only learn this way, you’ll end up missing a lot of non-obvious features of the language. E.g. writing your own function to do something you could do in a single standard library call, if only you knew it existed.

Something similar happens at one point in The Expanse (one of the best scifi series of the decade for sure).

Still dreaming of the day they use AI to enhance the street view experience. Being able to walk through a photorealistic simulation of any place on earth would be amazing, but the UX and image quality doesn’t seem to have changed in years.

You’re probably using the old Bard model. You can try the new one, bard-jan-24-gemini-pro, by clicking the Direct Chat tab on https://chat.lmsys.org.

What a fantastic site.

I was surprised to see that India appears to have a denser web of powerlines covering more of its area than the US and was wondering if this was a difference in data, but no, Wikipedia confirms:

> on average there is at least one HV line within a distance of 8.15 km) over the entire area of the country. This represents a total of almost 20% more HV transmission lines than that of the United States (322,000 km (200,000 mi) of 230 kV and above). However the Indian grid transmits far less electricity.

https://en.wikipedia.org/wiki/Electricity_sector_in_India#El...

It’s hard to believe ChatGPT would even work for this, unless this politician’s values and positions are identical to ones OpenAI have trained ChatGPT to have, the pretraining is going to leak through no matter what instructions they stuff into the prompt.