Gm. While there’s been a lot of AI newsletters popping up, we are not one. Although this edition, there’s some AI stuff we realllly need to tell you about.

🏠 ARTIFICIAL INTELLIGENCE

My New Job? OpenAI Safety Officer

If AI is the hottest field in tech right now, OpenAI is the Golden State Warriors. Since the launch of ChatGPT, Sam Altman and co. have racked up win after win:

• ChatGPT is the fastest-growing consumer app of all time.

• They received a $10B investment from Microsoft.

• They released GPT-4, by far the most advanced AI model in existence.

But, all this success has also come with criticism. Most of this criticism surrounds the lack of openness around GPT-4, but there is also a group critical of OpenAI’s ambitions in general.

Basically, these critics are afraid that OpenAI is leading us down a path of AI-fueled destruction.

So, to calm these fears, OpenAI has released their approach to AI safety. 

The 6 Keys To Safety

In their blog post, OpenAI specifically outlines 6 measures they are taking to ensure their AI models are safe:

1) Rigorous Testing: Prior to releasing any new model, OpenAI conducts “rigorous testing” to ensure that it’s safe for the public. For example, GPT-4 was tested and refined for 6 months before release.

2) Learn From Real-World Use: In the same way it’s impossible to predict all the ways AI will be used to improve the world, it is also impossible to predict all the ways AI will be abused. By tracking how their models are used in the real-world, OpenAI is able to improve things on the fly.

3) Protect Children: OpenAI requires you to be 18 or 13 with parental approval to use their models. They also have made it so that their tools can not be used to generate hateful, violent, or adult content, and have worked hard to ensure that the models won’t generate content that harms children.

4) Respecting Privacy: OpenAI has strived to train their models on public information, not personal data about private individuals.

5) Improving Factual Accuracy: OpenAI has used user feedback on ChatGPT outputs to improve its factual accuracy, leading to GPT-4 being 40% more likely to be correct than GPT-3.5.

6) Continued Research: AI is a new field with a lot of unknowns. Continued research is needed to ensure the safety of future more advanced models, and OpenAI is committed to being “increasingly cautious with the creation and deployment of more capable models”.

Why should I care? AI is awesome, but nobody fully understands the second-order effects yet. Extinction via AGI might not be in the cards, but other harmful consequences very well might be. We need the leaders in AI to make safety a priority, and it’s good to see OpenAI doing just that.

🏠 ARTIFICIAL INTELLIGENCE

Meta’s New AI Tool Named SAM

On Wednesday, Meta’s AI team released a tool named SAM (Segment Anything Model). Meta calls it a “step toward the first foundation model for image segmentation”. Yeah, because that means something.

So seriously what is Sam and why did over 2,000 people retweet the announcement tweet?

SAM is capable of one-click segmentation (identifying which image pixels belong to an object) of any object from any photo or video as well as zero-shot transfer to other segmentation tasks. Segmentation is crucial for computer vision and is used in a variety of apps ranging from analyzing scientific imagery to editing photos. If you’ve ever used photoshop, SAM powers stuff like the magic selector tool.

To democratize segmentation, Meta released the Segment Anything Model (SAM) and their Segment Anything 1-Billion Mask (SA-1B) dataset (the largest ever segmentation dataset) to There’s even a demo website where you can try out the model with your own photos.

Around the internet, people have been tweeting out pics of using SAM on brain imaging, retinal imaging, coral reef imaging, plant imaging, and more. “I can't imagine how big an impact this will cause to the community. Appreciate Meta's effort to release this to the public!,” wrote Wenhu Chen of Google Brain.

Why should I care? It feels like everyday there’s wild new improvements in AI. Most of the attention has been on LLMs for ChatGPT, but SAM will be a huge jump forward in imaging. From what it seems like on Twitter, anywhere AI imaging touches, SAM will have an impact.

🏃‍♀️ QUICKIES

What else is going on?

• iOS 17 is rumored to feature major changes to the iPhone’s control center.

• An Australian mayor is suing ChatGPT for defamation.

• Rocket-powered spaceplanes, coming soon.

• Apparently, it’s stupidly easy to open random garage doors. If you have a Nexx device, unplug it immediately.

Stat: 1.7x: Google claims its supercomputer is 1.7x faster than Nvidia’s. With AI becoming more and more prominent, the supercomputers used to train them will grow in importance as well.

Rabbit hole: Most Technologies Aren’t Races (Astral Codex Ten)

🤩 FUNDING FRIDAY

• Perplexity AI scores $26 million for their AI-powered search engine + launches their iOS app

• Honeycomb scores $50 million for their observability solution that shows you patterns and outliers in how users experience your code

• Chroma scores $18 million for their AI-native embedding database

• LayerZero labs raised $120 million for their omnichain interoperability protocol that enables the development of cross-chain applications

• Cranium raised $7 million for their app that makes sure companies’ AI is secure and trustworthy

🤖 The Time Machine

Remember the iMac G3?

Search “iMac G3 Aquarium” and you’ll see tons of people around the internet have turned their iMacs into fishtanks.

Funny + practical.

🛠️ FOUNDERS CORNER

The best resources we came across this weekend that will help you become a better founder, builder, or investor.

🧮 Quadratic - Work faster with your team via the first infinite pinch-and-zoom spreadsheet tool with Python, SQL, and Formulas.

💬 Build Your Own Telegram Chatbot - Wish there was a ChatGPT app? Follow these instructions to build a telegram chatbot with an OpenAI plug-in.

🐦 TweetHunter - Need a Twitter scheduler? TweetHunter is 🔥 and it even comes with AI writing and automatic retweeting of your best tweets.

💉 DOPAMINE HIT

Lol too real

@techroastshow

No such thing as a free lunch #techbro #techie #softwareengineer #siliconvalley #seattle #sanfrancisco #bayarea #comedy #startup #founder #unicorn

REACH 20K+ FOUNDERS, INVESTORS & OPERATORS

If you’re interested in advertising with us, send an email over to [email protected] with the subject “Homescreen Ads”.