- WeeklyDispatch.AI
- Posts
- Elon Musk live streams fully self driving Tesla
Elon Musk live streams fully self driving Tesla
Plus: WizardCoder-34B outperforms ChatGPT-3.5, Code Llama, and Claude-2 on HumanEval with 73.2% pass@1
The Dispatch
August 28, 2023
Welcome to The Dispatch! We are the newsletter that keeps you informed about AI. Each weekday, we scour the web to aggregate the many stories related to artificial intelligence; we pass along the news, useful resources, tools or services, technical analysis and exciting developments in open source. Even if you aren’t an engineer, we’ll keep you in touch with what’s going on under the hood in AI.
We hope your weekend was great! Today we’re covering:
Elon Musk attempts to silence Tesla critics
New ‘Moonwalker’ AI shoes help you walk three times faster
WizardCoder-34b trounces other coding LLMs on HumanEval
What Java devs think about the rise of GenAI
Trending tools and more
The Tesla CEO, along with Tesla’s Director of Autopilot Software, drove around Palo Alto for 45 minutes in Elon’s personal Model S. Image: Tesla Oracle
From Tesla Oracle: On Friday, Elon Musk live-streamed a 45-minute demo of Tesla's upcoming Full Self-Driving software version 12. This comes after years of Musk facing criticism for overpromising the capabilities and timeline of Tesla's self-driving technology. Despite bold claims that full autonomy was imminent, FSD has remained in limited beta for years. The pressure was on for Musk to finally demonstrate major progress, with this live test aimed at silencing the skeptics.
More details:
According to Musk, FSD 12 operates entirely on neural nets trained by video data, without predefined logic coding. Previously, Autopilot relied on hardcoded logic rules defined by engineers. Now, the system is entirely data-driven by neural networks, without any predefined code governing decisions.
FSD 12 drove smoothly through construction zones and made unprotected left turns, situations that challenged prior versions.
Musk revealed FSD 12 test drivers are active globally (to us, that means China), capturing training footage beyond North America to expand dataset diversity.
In the 45-minute drive, Musk had to intervene only once when the vehicle made an error at an advanced left-hand signal.
Takeaways: Near-perfect driving for 45 minutes cannot eclipse that one potentially dangerous mistake. Still, the progress is pretty stunning (and it’s not like the autonomous vehicles on the road now in San Francisco are doing better). The shift from logic programming to neural learning mimics evolution's move from rigid instincts to adaptable intelligence. If Tesla can continue improving at this pace, full autonomy may arrive sooner than critics expected.
Image: Shutterstock
From CNBC: Financial advisor Barry Glassman claims that current leaders in AI technology may not end up being the long-term winners for investors - but rather those that use it effectively in other industries. ChatGPT caused an explosive hype around AI, but the niche players today might not be the right long-term play.
More details:
Early internet companies like AOL and Cisco were incredible investments initially but later fell back down. A similar fate could await some of today's AI darlings.
There are few "pure play" public AI companies now. Leaders like Microsoft have AI as just one business line; chipmaker Nvidia has benefited from AI hype.
Companies in biotech, pharma, logistics etc. stand to gain immensely from leveraging AI - perhaps even more so than AI producers.
Takeaways: While it’s true that hype tends to outpace lasting value, the ‘top AI firms’ include some tech giants that - for now - appear to be poised to reap massive rewards from AI in the coming years. Still, C3.ai is one stock that certainly falls under this article’s umbrella - as would OpenAI if they ever went public. It seems hard to bet against the Googles and Microsofts of the world right now, though - even if you’re considering the long view.
 | Startup Shift Robotics has developed ‘Moonwalker’ devices that can be strapped onto shoes to increase walking speed by up to 250%. The AI-powered devices look like skates but move in sync with the user's movements, automatically regulating speed and allowing the user to lock/unlock them. A viral TikTok video demonstrates. |
After five years of development with robotics experts, the Moonwalkers rely on algorithms that adapt to each user's gait. They require no new skills (unlike rollerskating), can adapt to gaps and cracks in sidewalks of up to 2 inches, and seem aimed at city-walkers.
 | U.S. restrictions on investing in Chinese tech companies are having a major impact. Historically, American venture capital and private equity provided significant funding and expertise to help grow Chinese startups in AI, semiconductors, quantum computing and other key areas. New policies and rising geopolitical tensions have led U.S. investors to pull back. |
This is a major setback for China's ambitions to achieve self-sufficiency and become a global leader in these technologies. Chinese government funds are trying to fill the gap - but the government lacks the risk tolerance and long-term outlook of American capital. While Chinese firms are hyped as competitors, doubts remain about their viability without U.S. technical expertise and patient financing. The withdrawal of U.S. financial and technological support is becoming a huge challenge for China's AI/tech goals.
The professor’s great fear about AI? That it becomes the boss from hell
Arm’s IPO will tell us how much AI hype matters
Stephen King isn't afraid of AI - his books have trained it
(Opinion) ‘Generative inbreeding’ and its risk to human culture
Artificial intelligence seems poised to shake up long-held Catholic doctrine
From our sponsors:
The Wine Club for Real People
Bright Cellars believes that anyone can be a wine person. Learn what you like with our simple and fun wine taste quiz, and experience the joy of wine delivered to your door. No commitment. No fine print. Just wine your way.
WizardCoder-34B outperforms Code Llama, ChatGPT-3.5 and Claude-2 on HumanEval with 73.2% pass@1
Aside from the most recent tests with GPT-4, WizardCoder-34b is at the top of the HumanEval leaderboard (a harness for evaluating LLM’s trained on code)
The Story: While Meta’s recently released Code Llama is in the spotlight, open source model WizardCoder-34b has surprisingly achieved higher results on the HumanEval coding benchmark - supposedly surpassing even proprietary models other than GPT-4. How good is it, actually?
More Details:
Despite impressive metrics, no AI coding system today can write production-ready code out-of-the-box. LLMs are great at helping to solve constrained problems (but struggle with complex, multi-step challenges) - and WizardCoder is great at that. Extensive human guidance and code review is still required for real applications.
Reports can overinflate capabilities, and HumanEval isn’t a perfect evaluation system. Still, WizardCoder scored 73.2% - even topping GPT-4's initially reported 69.5% (more recent benchmarks have GPT-4 above 80%). From our testing, overall it is better than Code Llama and on par with Claude 2.
The model builds on WizardCoder’s earlier iterations and could reach GPT-4 performance levels in the future. But benchmark gaming remains an issue in the evaluation space.
Takeaways: WizardCoder performs better than any coding LLM outside of GPT-4 right now - although fine-tuned Code Llama models are popping up that might surpass it in the coming weeks. WizardCoder is developed by a team of Microsoft researchers.
Read more, and check out the demo and repos here. We’ve also linked a video breakdown from Prompt Engineering in the social media/videos section below.
 | Y Combinator's latest startup cohort contains a significant portion of AI-focused companies - especially those leveraging large language models (LLMs). They’re pursuing business models across the AI value chain: infrastructure, data/tooling platforms, and end applications. |
These AI startups are employing strategies like targeting niche customer problems, integrating with existing software, combining LLMs with computer vision and predictions, customizing models to user data, building creative interfaces, focusing on high information volume use cases needing precision, and enabling on-premise private model deployment. However, many may struggle to establish enduring competitive advantages versus large tech firms with the resources to replicate their offerings.
What do Java devs think of the rise of GenAI?
IBM’s brain-inspired analog chip aims to make AI more sustainable
What is actually open source has become a point of conflict around AI systems
Trending AI Tools & Services:
IndieZebra: A/B test your Product Hunt launch
Watto AI: ultimate writing companion for product managers
Lilac: an open source tool that enables AI practitioners to see and quantify their datasets
FineShare: explore a vast library of 100+ AI voice models and create song covers with your favorite AI vocals in just one click
GitBook AI: GitBook makes it easy to plan, centralize and share knowledge, from start to ship
Guides/useful/non-AI:
(Non-AI) Microsoft introduces Python in Excel
40+ cool AI tools you should check out (September 2023)
7 projects built with generative AI
(YouTube) How to fine-tune a ChatGPT 3.5 turbo model - step by step guide
How to use Google Bard for beginners
Social media/video/podcast:
Did you know?
The BRICS nations - Brazil, Russia, India, China, and South Africa - have agreed to form an AI ‘study group’. Chinese President Xi Jinping announced the decision at the 15th BRICS Summit on August 23rd, stressing the need for shared governance frameworks and standards to make AI more secure, reliable, and equitable globally.
The BRICS bloc represents 42% of the world's population and 23% of GDP, so cooperation between these major emerging economies on managing AI's opportunities and risks could have significant global impact.
It’s really smooth sailing in the car itself. It has never seen this construction before, but it’s just driving around it. There’s no line of code that tells the car to slow down for speed bumps or give clearance to bicyclists. This is all neural nets.