Miguel Jetté’s Post

I had the pleasure to chat with Dr. Joan Palmiter Bajorek about the work i've been doing for 8 years now at Rev! And it's true! We have a lot of folks working on our platform today, helping us do better and solve problems that make life a bit easier. It's always fun to think back of the road that led to where I am today. Working in AI and speech technologies has been a great adventure that started now 18 years ago. Tune in to learn a bit more about myself and Rev. Thanks for having me on Your AI Roadmap podcast! #ai #translation #transcription #subtitles #rev #humanintheloop #speech #speechrec #speechrecognition #asr #llm #tts

This is the type of announcement that I think doesn't get enough coverage. It relates to a post I just made before on how using powerful LLMs to solve "real problems" rather than toy problems. I guess it's not as "cool" and "flashy" as other types of applications. But providing accessibility to billions of people through translation is an incredibly important problem to solve. With this, Google is now powering 243 languages. It's an incredible achievement! Well on their way to provide translation services for 1,000 languages. https://lnkd.in/gvchd4Mu #ai #translation #subtitles #google

This sort of thing is great to see: https://lnkd.in/g4diGDqk Taking the right steps to access content the right way. It'll be interesting to see what this partnership yields over time. AI has the power to really revolutionize the way we interact with content and potentially find incredibly interesting nuggets of knowledge in this 101 years of humanity coverage. #AI #llm #trainingdata #openai

Just read this interesting article this morning: https://lnkd.in/g6tTwN9a I like it because it highlights a few problems with AI trends today: * Companies are rushing to integrate some of these AI models too quickly to "stay ahead of the buzz". * In doing these flashy demos, they create this false feeling that the models can do anything and everything under the sun. * Unfortunately, this behaviour erodes consumer confidence and sort of misses the mark on the power of some of these models. For some of these problems, it is fairly obvious that they are trying to use a sledgehammer to crack a nut. But these models are actually quite impressive for some tasks and can be used to really help solve serious problems. That's why I constantly like the use of these generative models as co-pilots to well defined tasks (i.e. keep the human involved!), and even using smaller purpose-built models to solve specific well defined problems (Like we do at Rev!). Of course, companies should make a huge effort to regain trust and confidence with their customers, they should always strive to publish results with the release of their models, and be upfront of where these models are used, how they are used, and why. #ai #responsibleai #trustai #rev

This recent study on Whisper's hallucination problems is super interesting (and somewhat concerning): https://lnkd.in/gAqRy95U > "in one sound bite, Whisper correctly transcribed a single, simple sentence, but then hallucinated five additional sentences that contained the words “terror,” “knife” and “killed,” none of which were in the original audio" At Rev, accuracy and reliability is non-negotiable. Unlike other transcription services that might ‘hallucinate’ extra sentences, we ensure every word transcribed is exactly what was spoken. #rev #ai #asr #speechtotext #stt #languagetechnologies #speech #whisper #openai

I like the insights in this paper review: https://lnkd.in/ggCejTFy The conclusion applies to more than linguistics. In an age where most people think that modeling is purely a data+GPU game, having linguistic knowledge can help a lot in understanding the problem at hand, explaining results, and testing properly. Especially in the language technology space. Of course data+GPU is most likely just as important (as we are seeing with these massive LLMs, MT, and ASR models), but having both will still give you an edge. #ai #datahungry #gpuhungry #linguistics #asr #mt #llm #machinelearning

Everybody's talking about GPT-4o and Google's Gemini releases this week... and most folks focus on some of the more "wow factor" features. I personally feel like most of the demos were sort of "smokes and mirrors" and mostly "cool marketing" demos. I'm not saying that the new models and their capabilities aren't impressive. They are incredible engineering feats. Training some of these models with all three modalities is certainly very impressive and super complex (I know audio features are complex to handle!). But these models are far from perfect, full-featured replacements to already existing products ... What they do is serve as a way to show what's possible, show what's coming, and help set a direction for the general public. Making them free and/or cheap to access and use means that they unlock possibilities for innovators try new ideas, iterate quickly, and find product-market fit on these ideas fast. 📈 The under appreciated problems (and main barriers to AI) remain: Bias In Bias Out, environmental impact (cost of training, etc), and Reliability. At Rev, we focus on smaller models that perform their tasks reliably, quickly, and in an unbiased manner. Let's continue to focus on solving real problems in an ethical manner. 💪 Having said this... I am glad to see Google include some of these AI features in their amazing products. I'm a huge fan of Google's products (gmail, Google Maps, Pixel Phone, etc). It'll be fun to see these features appear over time. #asr #ai #translation #bias #greentech

I just read this quick article on Bias in AI and how to tackle this problem. I definitely agree that many of the models out there are simply trained on "tons of data" that is not necessarily thoroughly vetted. The article suggests that the problem with Bias in AI will be tackled by understanding the data we feed the models, cleaning/filtering it, etc. In other words "Bias In, Bias Out". That's definitely something I like about how we are training our internal models at Rev. Our dataset is perfectly aligned with the problem we are trying to solve, thus directly vetted and filtered by the amazing people working on our platform. Bias in AI is one of the three biggest issues with AI today. And Sasha Luccioni presented this in a TED talk I shared recently (Look it up "AI is dangerous, but not for the reasons you think"): * Emitting carbon * Infringing copyrights * And spreading biased information Another thing I like to ponder with this... Are we even trying to solve the right problems with some of these GenAI models? Let's keep educating ourselves on the problems and do our best to leave things better than we found them. https://lnkd.in/g_hmP428 #biasinai #ted #bias #asr #ai #genai