Welcome back to SundAI, where the bots are evolving faster than you can hit “unsubscribe”. Get your popcorn, because this week was packed with over-the-top AI chaos. It also had just enough optimism to keep us all on edge.
Let’s roll.
If you like this article and you want to support me..
- Comment, or share the article; that will really help spread the word 🙌
- Connect with me on Linkedin 🙏
- Subscribe to TechTonic Shifts to get your daily dose of tech 📰
SundAI // Week 49 Waymo vs. Tesla: Who’s driving us to the future
(and who’s crashing along the way?)
Self-driving cars: humanity’s ultimate love-hate relationship. The question that is on everyone’s mind is if they will save us from ourselves or just create new ways to mess up our lives?
Tesla and Waymo showed off their wildly different approaches to getting your butt chauffeured by a computer this week.
Tesla’s FSD V13 is here.
That FSD stands for “Frighteningly Semi-Driverless”, for when you’re not sure if the car is driving you, or that you’re just along for a very, very terrifying ride….
Just kidding.. In Tesla’s vocabulary it means: “Full Self-Driving Version 12”.
Ok. Getting serious now. FSD allows your Tesla to drive itself from one parking spot to another anywhere in the US.
Sounds cool, that I must admit.
But if you were dreaming of sitting in the backseat, feet-up, drinking your daily bottle of whiskey and snorting a white line, you should not forget that there’s a little fine print saying: “driver supervision required”, because even mr Musk doesn’t trust his robots that much yet. Oh, and this model has 4.2x data scaling, 5x training compute scaling, and a 500% reduction in necessary driver interventions.
Not that this matters, because nobody apart from the product manager knows what the heck that is. Ok, let’s give it a shot….
Translation: fewer moments where you have to yell, “No, Tesla! Stop trying to murder that cyclist” (meaning, fewer errors and faster decisions)
And Waymo doesn’t want to be left out of the upgrade party, so it is rolling out 100,000 paid driverless trips weekly across three U.S. cities. And by doing just that, they are basically telling Muskus to stop dreaming, because they have proved it’s already living the autonomous dream Tesla is still chasing.
Their recipe for autonomy is a cocktail of 13 cameras (!), four lidars (distance and 3D mapping), and six radar units (for targeting pedestrians), all strapped onto their fleet of well ~700 cars. And meanwhile They’ve logged billions of miles in simulations and millions in the real world, all while Tesla’s over here flexing its 80 billion miles annually of fleet-collected data.
It basically means that Waymo is playing chess, and Tesla is playing poker.
Who is winning?
Well, that depends on your metric.
Waymo is a safer bet, but Tesla’s faster. Waymo is playing it safe with the FDA-approved shrooms of self-driving (carefully dosed and fully regulated), and while Tesla is on a coke bender, and it is tossing data at its models like it’s a wild Saturday night rave full of white lanes.
Alibaba releases an ‘open’ challenger to OpenAI’s o1 reasoning model
China’s Alibaba just spat Sam Altman right in the face with it’s QwQ-32B-Preview. That is the AI model which is designed to flex on benchmarks like GPQA and AIME. And while ChatGPT is as closed as a scared clam, Ali’s version of it is open-source under Apache 2.0, which means that anyone can use it for commercial shenanigans.
Take that, Sam!
Nous research is training an AI model using machines distributed across the internet
One of the folks at Nouse (boring website) probably woke up in the middle of the night, still halfway in their dream of creating a global scrapheap challenge by piecing together hardware from all over the world. That’s exactly what Nous Research is doing with their 15B parameter model. It’s like Frankenstein, but for AI, minus the maiden, torches and pitchforks… for now.
PRIME Intellect releases INTELLECT-1 – a global guerilla model
PRIME Intellect just materialized a decentralized large language turd, which is a 10-billion-parameter LLM cooked up across the globe like some kind of AI resistance movement. Forget the polished, expensive, energy-sucking, water-gobbling, planet f******g data centers, this model is basically a big middle finger to anyone who thought training LLMs needed a gazillion-dollar data center.
Their unholy algorithmic grail is the PRIME framework, which is built to laugh in the face of network unreliability and handles computing nodes popping in and out of the network like a sick game of whack-a-mole. It, though, is proof that decentralized AI is here to flip the bird to big tech’s centralized party!
ElevenLabs’ new feature is a NotebookLM competitor for creating GenAI podcasts
Initially, I left this news on the shelf, because I want you guys to think the podcasts at TechTonic Shifts were made by me, but since it’s out there now, I could not hide it anymore. So here we go:
ElevenLabs introduced GenFM, and that is a thingy for your AI-generated podcasts.
Think of it as Alexa hosting a talk show, but much, much creepier because it knows what you Googled last night.
You perv.
And I am not giving you a link…figure it out for yourself for once.
🖕
OpenAI’s Sora video generator appears to have leaked
And I do not mean that Sam has become incontinent, but that OpenAI can’t seem to catch a break. Because their unreleased Sora video generator leaked. And it was allegedly done by testers angry about, well, everything about ChatGPT – and their new GPT-PRO $200 offering. Who ever knew that AI ethics would involve so much workplace drama..
So rumor has it that Sora will be released during Xmas time (no, Elon did not steal Christmas). I just hope it will become part of their regular Premium offering instead of using it to promote the utterly overpriced, and underperforming Pro version.
AI2 releases new language models competitive with Meta’s Llama
Yeeeeeey.
Don’t pretend you even care!
The good thing about this is, that they are proving that open-source AI doesn’t have to suck.
If you are a frequent reader of my weekly AI-tabloid, you know that Llama is kinda open-sourcish. Meaning that it’s source is semi-open to only the popular people they invite to the party.
And now with A12, we are getting an open source model that is gunning for Meta’s Llama with the kind of quiet confidence that says, “We’re not here to play”.
Want to know why open source is the future of AI? Read: Open source is key for the future of AI | LinkedIn
Andrew Ng’s team releases ‘Aisuite’
The man with the simplest, yet most difficult name to pronounce: Andrew Ng and his band of brilliant misfits just made headlines because they made switching between LLMs as easy as changing your OF password. AI Suite is a New open source Python library for generative AI.
This means when you are sick of a particular model, you can simply change one parameter in your code and poof! A new LLM is live.
Uber for AI labeling
Because who better to label AI data than the same folks delivering your midnight sixpacks of Heineken.
If you are on LinkedIn, you will probably have come across ads from companies like Annotate, CloudFactory or Appen. These companies are willing to pay you big bucks (up to 4k through PayPal) for labeling data or just write some stuff so that AI companies can pay dearly to train their shiny new language models.
So Uber, with their horde of pedal-powered ninja’s have launched “Scaled Solutions”. And with that offering, it wants to turn gig workers into the backbone of AI development.
How much data would get labeled wrong if it’s a toss-up between Pizza, Burrito, or Noodles, because apparently, AI doesn’t do carbs.
What could possibly go wrong?
Short reads/videos to pretend you’re working
- Build a Chat-With-Document Application Using Python
- I Built an OpenAI-Style Swarm That Runs Entirely on My Laptop
- Faster Text Generation With Self-Speculative Decoding
- Multiphase Prompting
- Generative AI vs. Predictive AI: What’s the Difference?
Cool tools and some Git sh*t
- AI Suite: LLM juggling made easy.
- Langflow: Build your AI app dreams with minimal coding.
- Rerun: Visualize robotics data because spreadsheets are so last year.
- Keep: Alert management for the paranoid among us.
Exquisite research papers
- Unpacking DPO and PPO: Preference-based learning explained.
- MinerU: Turning PDFs into something actually useful.
- Mooncake: The architecture that makes LLMs go “wow.”
- DrugAgentAI for drug discovery. Breaking Bad, meet machine learning.
- SWIFT: Fine-tuning for people who want results now.
Random links
- Elon Musk’s Legal War With OpenAI: Popcorn-worthy drama.
- Meta’s SPDL Data-Loading Solution: Because AI training needs more acronyms.
That’s it for SundAI this week! Whether you’re training your basement AI or debating if Waymo or Tesla will kill us all first, remember: the future is coming, and it’s as weird as hell.
Signing off from the trenches of AI chaos, where burritos are noodles, and self-driving dreams ride on white lines,
Marco
Well, that’s a wrap for today. Tomorrow, I’ll have a fresh episode of TechTonic Shifts for you. If you enjoy my writing and want to support my work, feel free to buy me a coffee ♨️
Think a friend would enjoy this too? Share the newsletter and let them join the conversation. Linkedin appreciates your likes by making my articles available to more readers.
Leave a Reply