Gm. Sam Altman, you beautiful nerd, youโve done it again.

POLLS
We had a runaway winner for Wednesdayโs poll:

Something else: โA mirror, because you can reflect the sunlight off of it and the light will reach many miles in order to position yourself to be rescued. The chances are increased and thatโs the objective, to increase your chances of survivalโ Put me on a desert island with this guy.
Matches: โEverything you needโ: You sure about that?
Something else: โI would bring a bag with all the other items listed in it!โ: You clever clever dog.
Now for todayโs poll. Your answer will determine if we can be friends, so no pressure:
Who Is The NBA GOAT (Greatest Of All Time)?
๐ AI
The AI Training Debate

Training an AI model usually goes as follows:
The AI inhales an immense amount of data.
Humans double-check their answers to make sure they arenโt saying anything either incorrect (Sam Altman is the CEO of Wendyโs) or too crazy (hereโs how to kill your neighbor and get away with it).
This second step is known as โreinforcement learning through human feedbackโ, or RLHF for short. RLHF is how ChatGPT was trained, but itโs far from a perfect system:
Itโs time-consuming.
Itโs expensive (you have to pay the human trainers, albeit, not very much).
Itโs prone to human error.
These limitations are why Anthropic, the OpenAI competitor founded by ex-OpenAIโers looking to create a more ethical AI, is taking a different approach. RLHF is out, and Constitutional AI is in.
The basic idea of Constitutional AI is to imbue AI models with a set of values and then have the AI train itself. In practice, it looks something like this:
The AI answers a batch of questions and produces a set of first-draft answers.
The training model then assesses the answers and prompts the AI to rewrite the answers to be more ethical/non-toxic/helpful/any value you want.
The AI completes the rewrite.
The training model then trains the AI to write answers more like the rewritten version.
Itโs a pretty simple process, but one that Anthropic says is effective:
If Anthropic is right, and Constitutional AI is an improvement over RLHF, then itโs possible that we now have a way to train AI models quicker, cheaper, and more effectively.
The Big Picture: Considering Anthropicโs Constitutional advance, Googleโs AI coming out party at I/O, and the looming โno moatโ problem, OpenAI officially has some competition.
๐ NUCLEAR FUSION
Helion ๐ค Microsoft

Back in March, we wrote about Sam Altmanโs big bets on hard tech, specifically his $180M investment into longevity startup Retro and $375M into nuclear fusion startup Helion.
Longevity is cool, but thereโs a reason Altman put so much money into Helion. The world always needs more energy, and nuclear fusion isnโt just virtually limitless; itโs also clean. There isnโt anything like it, which is why Germanyโs decision to shut down its nuclear plants was so controversial.
What Is Fusion?
For the none nuclear scientists out there, fusion is when two light nuclei merge to create a single heavier nucleus. Because the new single nucleus is lighter than the two original nuclei, there is some leftover mass floating around. That mass becomes energy.
And not just any old energy, but an extremely powerful form of energy. Nuclear fusion powers the sun after all, and if itโs good enough for the sun, itโs good enough for us. Thereโs only one tinnnnyyy problem: itโs currently impossible to do.
Which is why Helion exists.
A Breakthrough?
And apparently, Helion has either made a major breakthrough or is confident a major breakthrough is coming soon, because they just signed a deal to provide Microsoft with fusion power by 2028.
Guess that means weโll have fusion power in the next five years?
Awesome!
Why should I care? Do you like having electricity? Do you like not living on an inhabitable burning rock? Then you should like fusion power.

๐โโ๏ธ QUICKIES
What else is going on?
Hardees and Carlโs Jrs are joining the AI drive-thru experiment.
YouTube has started blocking ad-blockers.
Microsoft doesnโt care about your annual raise.
Stat: 3 people: The first UK baby with DNA from 3 people was born after a groundbreaking IVF procedure. Modern medicine never ceases to amaze.
Rabbit hole: The Emptiness Of The Tech Mogul Myth (The New Yorker)
๐ค FUNDING FRIDAY
If you want a full list of startup raises including all the boring ones, you can go check Crunchbase. If you want the 2 coolest raises of the week, then check here every Friday.
๐ง AudioShake
AI music is having a moment right now, and AudioShake is capitalizing with a $2.7M seed round backed by Crush Music, Peermusic, and Metallica (lol). But AudioShake isnโt like the other girls AI music startups. Instead of trying to replace artists, AudioShake augments them.
What AudioShake does is allow for the breaking apart of songs into their respective components, like drums, guitars, and vocals, even if all of those components are only available on a single track. This allows for a whole new world of sampling possibilities, and the big three record labs are already on board.
Fire up the GarageBand fellas, weโre making music.
๐ง Pudgy Penguins
It must be 2021 because an NFT project actually raised money. This time itโs Pudgy Penguins cashing in with a $9M seed round. However, Pudgy Penguins has much larger ambitions than just NFTs. They envision the Penguins as toys, as clothing, and as characters in movies. Basically, as a powerhouse cultural brand.
A cool $9M definitely wonโt hurt those aspirations.
๐ ๏ธ FOUNDERS CORNER
The best resources we came across this weekend that will help you become a better founder, builder, or investor.
๐ซ StarChat is the AI coding assistant youโve always wanted.
๐ผ๏ธ Use Zmo.ai to create realistic backgrounds for your product photos. Itโs like your own personal green screen.
๐ HearHear is an AI listening bot that turns your meetings into actionable insights.

๐ DOPAMINE HIT
Lol facts.
HOW WAS TODAY'S NEWSLETTER?
REACH 35K+ FOUNDERS, INVESTORS & OPERATORS
If youโre interested in advertising with us, send an email over to [email protected] with the subject โHomescreen Adsโ.

