Agreed, this is VP Harris doing and saying the right thing. AI risks are not hypothetical, they're here right now, and people focusing on the hypothetical are either naïve or deliberately distracting us. mastodon.scot/@Peternimmo/1128…
This #Politico piece critical of Kamala #Harris at a recent #tech summit in #London actually put her up in my estimation. Rishi #Sunak and the tech bros wanted to talk about the fantasy "future #risks" of #AI.
Big model release today: Meta AI's Llama 3.1 series, including Llama 3.1 405B which appears to be the first openly licensed model that genuinely competes with current top proprietary models GPT-4o and Claude 3.5 Sonnet
We've been waiting for the largest release of the Llama 3 model for a few months, and now we're getting a whole new model family instead. Meta are calling Llama …
I'll be surprised if Llama 3.1 70B runs on your M2 with 64 GB of RAM. It doesn't run on my Windows PC with that much RAM, in Ollama, with the context window set to 128K. I don't have a discrete GPU, so host RAM versus VRAM isn't an issue. (And yes, running a model that big purely on CPU would probably be slow.) Apparently a big context window is a RAM hog.
This is a repository of apps to be used with F-Droid. Applications in this repository are official binaries built by the original application developers, taken from their resp. repositories (mostly Github, GitLab, Codeberg).
Der oder die FahrerIn eines dunklen SUV drängte am Samstagabend einen Radfahrer im Gegenverkehr auf Höhe einer Baustelle in der Paulinenhofstraße in Flittard ab. Infolgedessen stürzte der Radfahrer und erlitt mehrere Rippenbrüche. Der Verursacher flüchtete. presseportal.de/blaulicht/pm/1…
Köln (ots) - Am Samstagabend (20. Juli) gegen 20.30 Uhr ist ein 59 Jahre alter Radfahrer wegen eines mutmaßlichen Fahrmanövers eines dunklen SUV in Köln-Flittard gestürzt....
One writer's experience with AI summarizers: "ChatGPT isn’t summarising at all, it only looks like it. What it does is something else and that something else only becomes summarising in very specific circumstances." (Depending on the text, the training sets, and parameters, the writer notes it condensing the text without regard to its most salient points, recycling others' summaries of the same or similar texts, or fabricating details when asked for more information.) ea.rna.nl/2024/05/27/when-chat…
One of the use cases I thought was reasonable to expect from ChatGPT and Friends (LLMs) was summarising. It turns out I was wrong. What ChatGPT isn’t summarising at all, it only looks like it…
Cord cutting is continuing to hit Comcast hard as the company has recently lost both TV and internet customers. In Comcast’s 2nd quarter earnings report released this morning, the company reported it lost 419,000 cable TV customer losses.
Back in 2016 I wrote a book called Practical SVG. Recently, the publisher, A Book Apart, closed shop. Now you can Read Practical SVG on the web, here on this site, for free. I always like how Mat's book was online so now mine can join that cool club. I'll echo what I say on the site: Thanks to Jeff Eaton for…
Back in 2016 I wrote a book called Practical SVG. Recently, the publisher, A Book Apart, closed shop. Now you can Read Practical SVG on the web, here on this site, for free. I always like how Mat&#…
The first new Lyme disease vaccine in over two decades passed a key milestone last week — putting it on a timeline to potentially hit the market in 2026 if trial results are favorable. t.co/l0JIwiL64I
So thanks to @rmader@mastodon.social we managed to get the front camera in the Pixel 3a working with a libcamera/pipewire stack 🥳 Thanks to @flamingradian for the work on the whole port!
EDIT: The driver does not properly release the camera when closing the app, so it only works once per boot. It would be lovely if anybody with the skills would like to contribute to improve it!
Analogous to e.g. the imx219. This enables propagating V4L2_CID_CAMERA_SENSOR_ROTATION values so that libcamera can detect the correct rotation from the device tree and propagate it further to e.g....
> Or to put it another way: I actually view it as a red flag if an engineer or team gets into a predictable “flow” because it means that there is a promising opportunity for automation they’re ignoring.
Purism makes premium phones, laptops, mini PCs and servers running free software on PureOS. Purism products respect people's privacy and freedom while protecting their security.
I don’t think it’s appropriate to call the first Black woman presidential nominee by her first name. If you say Biden and Trump, you should say Harris.
Vorverurteilt und missverstanden – so fühlen sich viele Palästinenser, die ihre Stimme im Nahost-Konflikt erheben. Die Angst vor Konsequenzen ist groß.
"We explore the cost and security of reCAPTCHAv2 and conclude that it has an immense cost and no security. Overall, we believe that this study’s results prompt a natural conclusion: reCAPTCHAv2 and similar reCAPTCHA technology should be deprecated." arxiv.org/pdf/2311.10911
Ура! Подкаст, в котором мы записывались на прошлой неделе и с которым потом было так много странных историй, всё же выложен в архив вместе с кратким резюме. Мы с @Destranis были очень рады рассказать незрячим венграм о новом синтезаторе речи, с которыми в Венгрии прям беда и грусть.
Our final day of #GUADEC2024 BoFs and Workshops will start up soon. Make sure to check the website for BoF meeting links and your emails for access codes: events.gnome.org/event/209/pag…
Parabéns ao ganhador do Pants of the Thanks Awards deste ano, Felipe Borges @felipeborges ! Obrigado por todo o seu trabalho duro e muitas contribuições para o #GNOME! 👣 fixupx.com/guadec/status/18150…
MacOS kernelové ovladače třetích stran zakázal, v Linuxu běží Falcon jako eBPF. Za to, že Microsoft nebyl schopný poskytnout vhodná API, není přece problém antimonopolní regulace.
A ta perlička o úderu blesku - Kurtz to posral v McAfee a teď to podělal v CrowdStrike. To, že se podobná chyba stala na dvou místech zároveň je záležitost firemní kultury, viz např. zde: hachyderm.io/@shanselman/11281…
V pátek se zastavil provoz v největším evropském přístavu v Rotterdamu, na zemi zůstala letadla desítek leteckých společností. Jen v USA došlo během pátku a soboty ke zrušení více než pěti tisíc letů.
Že to hlásá Microsoft, který využívá situace, chápu. Ale, že ten blábol přebírají Hospodářky. 🤦♂️ To je jak tvrdit že za špatné jídlo v restauraci může hygienická stanice, která kuchaře obtěžovala s čištěním kuchyně.
No, Southwest Airlines is not still using Windows 3.1
A story that's been persistently making the rounds since the CrowdStrike event is that while several airline companies were affected in one way or another, Southwest Airlines escaped the mayhem because they were still using windows 3.1. It's a great story that fits the current zeitgeist about technology and its role in society, un
To je snůška kecu jak od nějakého dezolata, ano zla EU může za to že Microsoft neumí udělat APi tak aby to nepodělalo ten jejich shit který nazývá prý operačním systémem ..m
Koncem předminulého desetiletí hrozilo americké firmě vyšetřování a tučné pokuty ze strany Evropské unie, a to kvůli údajně nesoutěžnímu jednání. Takže roku 2009 raději s Bruselem uzavřela dohodu
Oh, the new option resolver-use-dns64 is cool! 👍🏻 bind9.readthedocs.io/en/v9.20.… That means you can run your DNS64 resolver as a part of your IPv6-only infrastructure without provisioning it with dual stack connectivity or CLAT.
It's vital, over the next few weeks, to remember that Republicans, foreign powers, and extremists will be working to make Democrats and people who vote for them angry with each other.
In dictatorial countries, it's best to use cloud services for both files and passwords so that you have as little sensitive information on your devices as possible
Recommending password-manager? Good idea! Rcommending 2FA/MFA? Even better! Recommending to put both password and second factor in the same password-manager? Meh... I don't know, if you meant it that way, but it somewhat reads it ...
@cart1082 Thanks for the feedback! We would agree that it is best not to keep your eggs in one basket. Secure unique passwords stored in a password manager + a physical U2F key is our go-to wherever possible.
This post was a bit surprising: Mozilla's ad tracking feature is part of an "origin trial", which means the new API is only exposed to websites on an allow list. Currently that list is empty, so the it doesn't do anything even if you've left the feature enabled.
They generated all that bad publicity, and the feature isn't even collecting any data yet.
that's cool and all, but isn't this what libreelec is basically made for? That's what I use to control my tv amd watch netflix or whatever streaming service.
I'm a big fan of LibreElec, have a RPi4 running it for my TV. But it's very much Kodi and no other apps. Nothing for viewing PDFs, no browser, no messaging, just Kodi. Which is generally fine! But this hack-a-TV can do all of those other things, and has more CPU grunt for it.
I'm glad to share with you this early sample of my new neural TTS model OptiSpeech. The model is end-to-end, and this variant is based on ConvNeXtTTS architecture. Noticeably, the model is trained using 24KHz sampling rate, while being very performant. Thanks for Pneuma Solutions for providing access to the hardware I'm using to train the model.
@matt The dataset used for training this voice is released under a non-commercial license (CC BY-NC-SA 4.0). Here's the release announcement page: ast-astrec.nict.go.jp/en/relea…
Thanks. Also, is optispeech still using the espeak phonemizer? Or when you say it's end-to-end, do you mean the model goes directly from text to speech?
@matt It does use eSpeak-ng's phonemizer since a good and portable phonemizer is not currently available. But the developer of Piper is currently working on an MIT licensed phonemizer that's compatible with eSpeak-ng's phoneme-set.
@matt A truly end-to-end TTS model, that maps from characters to waveforms, is not currently possible except with extremely large model sizes. It can be done with compact models, but don't expect good results with phonemically irregular languages like English.
Is there any data yet on how many utterances the training dataset needs to have? I assume there's a reason why the HFC datasets have as many utterances as they do, but I wonder if that decision was based on the needs of models like yours, or something else.
@matt Regarding single speaker datasets... There is no exact figure. Not scene a paper analyzing that neither. The general consensus is that at least 10 hours is needed for good quality speech. 30 to 40 hours seams to be the norm in the industry. But it also depends on the model. Some models are data efficient; learning from very few utterances. What people usually do is to train on a single, high-quality and large dataset, then fine tune the model on other smaller datasets.
@asael @matt There are multiple good data sets available. You can check out Piper's voices repository on hugging face. With each voice comes a model card that contains information about the data set used to train that voice.
@asael I found and even easier way for you. 1. Visit Piper samples page rhasspy.github.io/piper-sample… 2. Select the language/voice you want to learn about 3. After the audio player you'll find a read-only edit box containing voice information, arrow up to the dataset section. Note that some datasets are exclusive, but most of them are openly available.
@Lino_JR I chose it because all the good names have been taken by other TTS models 😄 Fast Speech, Light Speech, Lean Speech, Efficient Speech, Speedy Speech..etc.
How does it handle punctuation differences? My biggest pet peeve with virtually all synths apart from Espeak and Eloquence is that a comma, period, exclamation point, semicolon, two dashes etc. all sound the same! Same amount of pause and same voice inflection. With eSpeak, for example, the voice will pitch up and sound more excited if the sentence ends with an exclamation point. Is that something LLM’s can do?
@drew It's one of my biggest issues too. And it's one of the things I'm trying to address through this effort. It largely depends on the model's design, given you have a good dataset. With large models, it is almost a solved issue, but with compact models it is a challenge. Large models have grater capacity to learn sentence-level as well as character-level prosody.
hey @FreakyFwoof. For some reason I had a copy of this 3 years ago that I don't remember reading.
Finished it yesterday. it was ... kinda OK, guess. clever twist on if just one small thing had been different.
A Better Man By: Valandar Vernon changed in the beginning to be a better man, thanks to a wish. How will this affect Harry, and the entire Wizarding World? Now complete! Please R&R. AU, HPGW, RWHG, RLNT Rated: Fiction T - English - Drama - Ginny W., Harry P. - Chapters: 54 - Words: 192,896 - [A Better Man](fanfiction.net/s/2531438/1/A-B…)
Neural networks surround us, in the form of large language models, speech transcription systems, molecular discovery algorithms, robotics, and much more.
![congratulations.
you've scrolled so far you found the dopamine!
[drawing of the molecular configuration of dopamine]](https://fedi.ml/photo/preview/640/574252 "congratulations.
you've scrolled so far you found the dopamine!
[drawing of the molecular configuration of dopamine]")
Matt Campbell
in reply to Simon Willison • • •