🏠 Neurons, Networks, and New Knowings
Mechanistic Interpretability, You've Been Solved
Gm. Last Friday I asked for some NY Tech Week events, and now I have one to share with you all. Homescreen is co-hosting a party on Wednesday, October 18th. If you’re looking to mingle with some tech compatriots, RSVP here.
BROUGHT TO YOU BY:
Anthropic was founded as a responsible AI startup. Basically, they want to create cool AI products, but what they really want is to make sure AI is aligned with humanity’s best interests.
In English, we understand the math behind neural networks (AI models), but we don’t understand why the math makes neural networks do the things we see. This makes it difficult to both advance onwards and certify that a model is safe. After all, how do you diagnose something you don’t even truly understand?
If this sounds similar to the problem neuroscientists face with understanding how neurons affect human behavior, well, that’s because it’s the same problem. Except, thankfully, AI researchers can run experiments much easier. And that’s exactly what Anthropic did.
Through these tests, they found that individual neurons do not come with individual roles. For example, the same neuron can play a part in academic citations, English dialogue, and Korean text. That’s a problem, but Anthropic found a clever fix. Instead of examining individual neurons, they built machinery that bunches up similar neurons into units called features.
The result is that we now have the ability to understand exactly why a model is behaving the way it is, and, perhaps more importantly, the ability to change the model’s behavior.
Does this mean that an aligned superintelligence is now guaranteed? Not quite.
But does it mean that we took a large step forward in the quest for responsible AI? Definitely.
The Big Picture: A safe and controllable AI seems to be rapidly approaching.
Mobile apps = iPhone. AI apps = Wispr
Wispr is building a wearable brain-computer interface that lets you communicate at the speed of thought. This is the next leap in personal computing – a consumer hardware platform for AI-first applications.
While AI has evolved with LLMs, we still interact with them with interfaces from the 90s - primitive textboxes and chat interfaces. To redesign input interfaces, Wispr has raised close to 15M from NEA and 8VC in addition to Stanford AI professors like Chris Manning & Richard Socher’s fund AIX ventures.
A big revolution is coming. And it’s going to be spearheaded by innovation throughout the stack.
Wispr is recognized as one of the top 100 AI companies by CBInsights – top 4 in the “AI assistants & HMIs” category, along with Anthropic, Inflection, and Adept AI.
🤝 THE LATEST IN…
Raise: Osmoses raised a $11M seed round to bring its molecular-scale membranes to the hydrogen market.
Stat: $3-4.5B: How much FTX’s stake in Anthropic could now be worth, which is enough to make the exchange whole. The SBF hail mary succeeds!
Rabbit hole: Making Hard Things Easy (Julia Evans)
I actually am in the minority here and think Google’s going to win this case, but we’ll see:
Let’s stay on the monopoly theme this week:
Will Amazon lose its monopoly case?
🛠️ FOUNDERS CORNER
The best resources we came across this weekend that will help you become a better founder, builder, or investor.
🎙️ Looking to raise soon? You need this workshop.
🔥 Pitch your startup idea, but be warned, you will be roasted.
💰 If you’re an investor in need of some new deal flow, this is for you.
💉 DOPAMINE HIT
Tobey Maguire sighting!
HOW WAS TODAY'S NEWSLETTER?
REACH 40K+ FOUNDERS, INVESTORS & OPERATORS
If you’re interested in advertising with us, send an email over to [email protected] with the subject “Homescreen Ads”.