Sumeet Maniar’s Post

Really interesting read on the Tesla FSD vs. Waymo approach. There are challenges for both. The way l see it, Tesla is trying to solve a software problem through scale. Theory goes more data equals better performance (simply put). Waymo has a software, hardware and scale problem. Tesla has a hardware challenges too, of course, but it may be they will better the type and weatherproof vs expanding their hardware set. I'm not quite convinced Waymo is better with edge cases than FSD. It would be great if one was devised, with everyday drivers devising the cases. I vote for cab and truck/delivery drivers since they've probably seen it all. And yes ALL AI systems take a best guess approach at correctness when it comes to the less familiar. But quite honestly so do too many drivers on the road. https://lnkd.in/evvmxHhd #fsd #waymo #ai

Google's Waymo was caught driving on the wrong side of the road. No one was injured, but it would be a scary sight seeing a car trying to drive. There was a passenger in the driverless taxi. This video is taken out of context a bit. How did the car end up in this position? Waymo cars use a neural network called VectorNet. Like all neural networks, it uses probability to determine the best outcome. In this case, the car determined the best path was quickly veering into oncoming traffic. It is very likely the car was either confused and/or had multiple inputs that conflicted with each other. It could also be likely the car simply had no other option. It had to make the best bad decision to get out of a bad situation. Neural Networks are never going to be perfect. There are always going to be 'weird' situations that trip up AI models. Humans aren't perfect either. Driving is a very imperfect activity that requires a lot of decision making and judgement calls. In this case, without seeing the full scope of what happened, I think the car did ok. No accidents, no one was injured, no cars were damaged. I see the hazard lights are on, the weather is bad, and traffic was a bit of a mess anyway. There are 36,000 accidents every day in the United States. Whatever this car did, it didn't get into an accident.

#Wayve is emerging as the next AI decacorn. They just raised an absolutely massive round, over $1billion series C led by Softbank, Nvidia, and Microsoft, the biggest round ever in the U.K. Their VLAMs - Vision Language Action Models - are a breakthrough in #neurosymbolic computing and #explainableAI - natural language for autonomous driving - feeding data from a vehicle's cameras and actions into driving manuals and a road's geographic rules, into a language model to provide a continuous commentary and explainability. The natural language commentary could also become very high quality data to feed into #syntheticdata generation models. combining driving actions and vision models with with natural language.

What if LLMs were the unexpected answer to autonomous driving? Jérémy Cohen dives into this topic in our latest article, covering topics such as HiLM-D, MTD-GPT, PromptTrack, MagicDrive, and many more! Read it here: 👇 https://lnkd.in/grBAFw78

This new Andrew Hawkins piece from The Verge in anticipation of the Tesla “We, Robot” Robotaxi event covers the state of the art of driving automation, including #autonowashing. 📺 Will you be watching #werobot? What do you expect from it? 🔮 What ratio of hype vs. reality do you predict? Thanks to Jim Gibbs for bringing this article to my attention 🙌 https://lnkd.in/exh9MkH4 #automatedvehicles #AI #automation

Waymo Builds A Vision Based End-To-End Driving Model, Like Tesla/Wayve #DL #AI #ML #DeepLearning #ArtificialIntelligence #MachineLearning #ComputerVision #AutonomousVehicles #NeuroMorphic #Robotics

End-to-end” is what Tesla refers to as neural net-powered AI driving the vehicle from vision to controls rather than the controls being explicitly coded. It’s already the case in all widely released versions of (Supervised) Full Self-Driving (FSD) for city driving, but not for highway driving, which uses another software stack.

6 LEVELS OF AUTONOMOUS WORK: Digital Assistant — Work Augmentation LEVEL 1: Task LEVEL 2: Sub-Process LEVEL 3: Process (2024) Digital Agent — Work Replacement LEVEL 4: Role (2025) LEVEL 5: Team (2030) LEVEL 6: Line-of-Business/Org (2035) We used the Tesla Robotaxi economy and the autonomous vehicle evolution as the reference architecture. ZDNET https://lnkd.in/eFKdKpDK

OK, so is this a sign that true self-driving cars will remain a fantasy? Personally I've always struggled seeing how a probabilistic approach could ever deliver self-driving cars - there will always be a near infinite long tail of events with little or no data, ensuring that a reasonable proportion of choices will not be able to be handled by statistical evaluation. For self-driving cars to really work we need proper problem solving intelligence, something we currently just don't have. The benefit of focussing on generative text & image AI is that people don't die if the probabilistic approach fails, and also it's great for hype generation (not that I'm being cynical ;-))

Our very own Steve Lowry was interviewed in the latest article by Sharp Magazine 'Welcome to the Future'! Dive into the discussion about the transformative power of AI and the innovative trends shaping our future. Steve shares insights on how AI is seamlessly integrating into our daily lives. As he puts it, “The new DJ feature on Spotify is simple yet quite helpful and fun. It’s one less thing to think about — a good jam to put you in the right mood.” He also highlights the future of AI in driving, noting, “In perhaps five years, autonomous driving will start to become commonplace... In the long run, computer drivers will be far safer than human drivers.” Discover how AI is making life more effortless and enjoyable. Check out the full article here: https://lnkd.in/g_HpbXms