- Superhuman AI
- Posts
- The limits of using the web to train AI
The limits of using the web to train AI
ALSO: Create a custom analytics dashboard with Claude
Zain Kahn
June 25, 2024
Read time: under 4 minutes
Welcome back, Superhuman
Most semiconductor companies design AI chips that can do lots of different things — since it’s hard to predict which algorithm will be most popular by the time a chip finally hits the market years later. In today’s newsletter, find out why a startup called Etched took the exact opposite approach.
Today’s Insights
The limits of training AI on the internet
Tutorial: Build an analytics dashboard with Claude 3.5
Chart: The first algorithm-specific chip
5 new AI tools to boost your productivity
Everything else you should know today
AI-Generated Images: Marble Woman
NEXT IN AI
A debate over how much AI companies should scrape from the internet
For decades, web pages have used a piece of code called robots.txt that acts like a museum security guard. It says “You can look at our site all you want, but please don’t take anything.” Well, according to a content licensing startup called TollBit, AI companies have lately been sneaking in after hours and grabbing things right off the walls.
What’s the evidence? Tech publications Wired and Forbes recently claimed that some of their articles had been lifted and republished on AI platform Perplexity without proper attribution. But TollBit emphasized that the problem is bigger than any single company. OpenAI, Anthropic, and others also appear to be ignoring these longstanding conventions. “The more publisher logs we ingest, the more this pattern emerges,” TollBit wrote.
What’s the defense? Some AI companies claim they’re simply referring to other sites’ content, not just copying it wholesale — sort of like showing off a photo of the art inside a museum, while encouraging people to go see it for themselves in person. For its part, OpenAI told Business Insider it takes web permission “into account each time we train a new model.”
What’s to come? A lot of it will pivot on how Congress and the Supreme Court interpret AI-generated content. Humans incorporate the knowledge they’ve gained from books, articles, and other material into their work all the time. But you can’t just copy something word-for-word and pass it off as your own. In order to stay in line with copyright law, AI companies will likely have to prove that their models are actually creating something new, not just rearranging the content they’ve already been exposed to.
PRESENTED BY GUIDDE
Create video documentation 11x faster with AI
Tired of explaining the same thing over and over again to your colleagues? It’s time to delegate that work to AI. Guidde is a GPT-powered tool that helps you explain the most complex tasks in seconds with AI-generated documentation.
Turn boring documentation into stunning visual guides
Save valuable time by creating video documentation 11x faster
Share or embed your guide anywhere for your team to see
Simply click capture on our browser extension and the app will automatically generate step-by-step video guides complete with visuals, voiceover, and calls to action.
The best part? Our extension is 100% free.
AI AT WORK
How to create an analytics dashboard using Claude 3.5
Go to Claude’s website and log in with your account.
Go to your profile and select Feature Preview.
Toggle the Artifacts button.
Now, upload your document with data and ask Claude to turn it into an analytics dashboard.
Use the prompt:
“Make an analytics dashboard tracking market campaign performance”
Wait for a few seconds and you’ll have your analytics dashboard to track performance.
You can preview the dashboard using the Artifacts feature or you can also preview the code to be used in the future.
PROMPT OF THE DAY
Self-help
Prompt: I want you to act as a self-help book. You will provide me with advice and tips on how to improve certain areas of my life, such as relationships, career development, or financial planning. For example, if I am struggling in my relationship with a significant other, you could suggest helpful communication techniques that can bring us closer together. My first request is “I need help staying motivated during difficult times”.
You can adapt the prompt to your specific needs.
Source: Beebom
PRESENTED BY TENSORWAVE
Finally, a cloud provider built for heavy AI workloads
Don’t let wait times kill your productivity. TensorWave’s cloud leverages AMD’s Instinct™ MI300X accelerators, designed to deliver high performance at scale for gen AI workloads and HPC applications.
CHART
The world’s first transformer-specific chip
Source: Etched
In 2022, before the launch of ChatGPT, a startup called Etched was already convinced that transformers would one day become the algorithm-of-choice for AI companies. Transformers — the neural networks that power GPT-4, Llama, Stable Diffusion, and nearly every other major LLM — are unique because they can break sentences down into individual words and figure out how to use those words in future contexts.
Most GPUs can handle lots of different AI-oriented tasks. But Etched says it’s built the “world’s first transformer-specific AI chip” — one that was designed from the ground up to work exclusively with transformer algorithms, and nothing else.
Etched claims that while competitors like Nvidia have been building chips that get bigger and bigger, they haven’t managed to develop ones that are also more powerful — at least not in the past few years. That’s because most AI chip companies have to save space on their semiconductors for algorithms besides transformers.
The San Francisco-based startup’s new Sohu chip, meanwhile, is allegedly magnitudes faster and more efficient than its rivals, including Nvidia’s state-of-the-art H100s, because all of its space is dedicated to just one task. One chart shows that Sohu helped Meta’s latest Llama-3 model run 20 times faster than competitors’ chips.
The company thinks AI startups will come to rely on its chips as they continue to scale up toward super-intelligence — adding that it has already courted tens of millions of hardware reservations from early customers. It recently raised $120 million in a Series A funding round, including contributions from Peter Thiel, Kyle Vogt, and other notable investors.
PRODUCTIVITY
5 AI Tools to Supercharge Your Productivity
✅ Spiral: Automate 80% of repeat writing, thinking, and creative tasks, while maintaining your unique voice and style.
✅ Astrocade AI: The first platform that makes building games and interactive experiences easy enough for creators to share their ideas in minutes.
✅ Thryve: An AI-powered companion for emotional and wellbeing support across 50 languages.
✅ Hoppy Copy: Use AI to create powerful content for hundreds of different email campaigns, newsletters, and more in seconds.
✅ Remodel AI: Take photos of your home and instantly see a fully remodeled version, with new flooring, different walls, and more.
PS: Want more? Check out our Top 100 AI Tools.
* indicates a promoted tool, if any
AI & TECH NEWS
Everything else you need to know today
Source: Getty Images
Unlikely Partners: TikTok’s parent company ByteDance announced it’s collaborating with California-based Broadcom on new AI chips, despite the US’s attempts to cut off Chinese companies from state-of-the-art semiconductors.
New Frontiers: In its first major expansion into the Middle East, Nvidia plans to supply telecom giant Ooredoo with powerful GPUs across data centers in Qatar, Algeria, and at least four other countries.
Double the Trouble: The European Commission claims Apple breached its Digital Markets Act by blocking competition to its app store. That came just days after the iPhone maker said it would delay the rollout of AI features in Europe due to stringent regulations.
Silicon Valley Says No: Venture capitalist firm YCombinator joined around 140 startup founders to condemn a California bill that would require AI companies to conduct risk assessments before releasing new models.
Lawsuit Blues: The “big three” music labels are suing Suno and Udio — two popular text-to-song generators — over claims that both AI companies trained their models with copyrighted songs and albums.
😄 One Fun Thing: A pair of best friends who just graduated from high school in Colorado are already developing their own AI company. Their startup is called APIGen, referring to the applications that help different pieces of software communicate with each other. Clients will be able to create new APIs from scratch with natural language — a task that used to require advanced coding expertise.
AI-GENERATED IMAGES
Marble & Coral
Source: @r_iii on Midjourney
Prompt: A marble statue of an ancient Greek woman, sitting on the floor with her head covered in a white linen cloth and holding a small plate of food. She is surrounded by coral plants and sea foam, in the style of an ancient Greek artist. --ar 3:4
Acquire new customers and drive revenue by partnering with us
Superhuman is the world’s biggest AI newsletter for businesses and professionals with 600,000+ readers working at the world’s leading startups and enterprises. Companies like Amazon, Hubspot, and Salesforce feature their products in Superhuman. You can learn more about partnering with us here.
