😺 Nuclear power for ChatGPT?

Welcome, humans.

GPT users can't help themselves—riffing off of yesterday's trend, now everybody’s asking ChatGPT which fictional characters they most resemble based on their chat history:

Looks like ChatGPT is creating an army of Hermione Grangers, Tony Starks, and Jean-Luc Picards (not sure how creative GPT’s being here).

We’re just waiting for that one person brave enough to admit they got Jar Jar Binks…

Here’s what you need to know about AI today:

• Google joined the big tech leaders using nuclear to power AI.
• Microsoft's VP of GenAI research joined OpenAI to work on AGI.
• More GPT users gained access to searchGPT, iOS app update expected.
• Rumors of new ChatGPT model circulated on social media.

Advertise in The Neuron here.

Let’s talk about AI and energy… nuclear energy.

AI uses A LOT of energy—it’s been estimated that training GPT-3 would take roughly the same power used by ~130 homes for a whole year.

That’s just training. Using AI on the daily:

• A ChatGPT query = 10x more energy than a Google search.
• AI-powered email = a bottle’s worth of water to cool the servers.
• Running a small model for 1M people = powering 5 homes for a whole year.

Perplexity with the why.

Big tech's solution = Nuclear power. Example: Google's teaming up with Kairos Power to build 7 small nuclear reactors for its data centers by 2030.

But here’s the thing: these are a new type of nuclear reactor that are largely untested, with economics that don’t necessarily make sense—or cents.

Think that’s optimistic? Listen to what everyone else is doing:

• Amazon built a new $650M data center next to a 2.5 gigawatt nuclear plant in Pennsylvania that’s now stuck in regulatory limbo.
• OpenAI wants to deploy multiple 5 gigawatt data centers in different cities across the U.S.
• Microsoft signed a deal with a different Sam Altman backed startup for 50 megawatts of fusion power (largely considered the holy grail of energy) that’s ”expected” to come online in 2028. What, like it’s hard?

Microsoft’s so desperate for power, it signed a deal to reopen Three Mile Island—y’know, the nuclear reactor famous for being America's favorite nuclear oopsie-daisy?

Here’s why: the AI companies are getting cut off.

1. Salt Lake City and Santa Clara put a moratorium on “large power requests.”
2. Virginia, dubbed “data center alley”, is temporarily rationing power to new builds.
3. Ohio's AEP plans to charge data centers more, anticipating “Three New Yorks” worth of power demand by 2028.

But AI's computational power needs double every 100 days. A 10x efficiency improvement could mean 10,000x more power demand. By 2026, data centers could consume as much electricity per year as the entire country of Japan.

Source: SemiAnalysis

There might be a silver lining: a new computation method claims to cut AI’s energy needs by 95%—no fusion required.

FROM OUR PARTNERS

You can now unlock instant insights from any voice conversation with Gladia Real-Time.

Your customer support and sales teams need actionable, real-time insights to boost CX and close deals.

Well, Gladia just made that dream come true:

• Gladia’s new Real-Time API is like having an assistant feeding you real-time insights at your fingertips at all times.  
• Its real-time analysis detects key entities in a call, and even understands caller sentiment.
• It transcribes speech in 100+ languages with barely any delay (we're talking 300 milliseconds) even if you switch languages–and it's scary accurate.

It also plays nice with all the usual stacks and phone systems (SIP, VoIP, FreeSwitch, Asterisk), so it's a breeze to set up.

Oh, and did we mention they just raised $16M to keep pushing the boundaries of AI audio tech? Yeah, Gladia means business.

Click here to be the first to try.

Around the Horn.

Open source Apple image models. Code here, paper here.

• Microsoft’s VP of GenAI research (Sebastien Bubeck) has joined OpenAI to “further his work toward developing AGI.”
• Someone on X shared a screenshot from ChatGPT that references a model “gpt-4.5o”, which has some people thinking a new model is dropping soon—but from the comments, there’s reasons to doubt, so take this with a grain of salt.
• In related news, more users are starting to get access to searchGPT, and it looks like the iOS app version may get it in a new update.
• Google’s AI Studio may soon get its own Canvas feature like ChatGPT, and AI Studio will start grounding its responses with facts from trusted sources.
• If you’re interested in OpenAI’s journey from non-profit to a $157B corporation, these tax documents show how it started. Also, apparently OpenAI is in a lawsuit with… Open AI (with a space).

Treats To Try.

1. *Sana AI is your ultimate AI assistant for work—it helps analyze documents, summarize meetings, automate repetitive manual tasks, and more. Work faster AND smarter—try Sana AI for free today!
2. Adobe announced new AI features for Photoshop, including Distraction Removal that lets you erase unwanted objects (like that gawking rando) from your images, and more that are now widely available in Creative Cloud.
3. Piclooks gives you a library of authentic looking generated profile pictures to download for free, saving you time on design projects.
4. Omnic.AI analyzes your video gameplay footage to provide personalized tips and statistics to help you improve your performance.
5. Anam lets you add lifelike digital humans to your product that can chat with customers in 32 languages (beta atm, try it here).
6. LLMWare.AI gives you tools to build AI apps for regulated industries like finance and law (new product launch, read more here).
7. Zamba2-7B is a newly-launched open-source model in collaboration with NVIDIA, and its supposedly one of the best small models available (Hugging face, API details)—try it out yourself here.

See our top 51 AI Tools for Business here!

*This is sponsored content. Advertise in The Neuron here.

NEW SECTION: Prompt tip of the week.

Have you noticed we haven’t given out much prompt advice lately? That’s because with ChatGPT o1, writing a long detailed prompt is a lot less important than it once was.

However, that doesn’t mean prompting with o1 is easy; here’s some advice we recommend you try this week:

1. Before prompting o1-mini or o1-preview, pull up 4o to help write your prompt.
2. Ask 4o something like this:
   1. “I'm writing a prompt for the o1 model and I want to make sure I haven't overlooked anything before submitting. Take the following prompt, and ask me the most important question that could further improve it. For the output, give me the prompt again. Continue the prompt improvement questioning feedback loop until I say so.”
3. When you’re satisfied, take this prompt to o1-mini first (50 prompts per day) and go to o1-preview only when mini can’t solve your problem.

The best way to think of ChatGPT’s model stack rn:

• 4o = I’m feeling chatty. Vent, plan, brainstorm, etc.
• o1 mini = I want a specific task done.
• o1-preview = for when you hit a wall and need to pull out the big guns.

A Cat's Commentary.