🧠 Google solved AI's long-term memory

Google just unveiled what it sees as the successor to the Transformer. It is called Titans, a model that can memorize while it is running. Many researchers are already asking the same question. Is this the breakthrough everyone has been waiting for? Let’s dive in!

In today’s insights:

LATEST DEVELOPMENTS

Evolving AI: Google has introduced a new architecture called Titans. It gives models real long-term memory during inference, which could change how large models handle context.

Google has unveiled Titans, a new architecture that lets models update their long-term memory while processing text. Instead of relying only on fixed weights, Titans uses a deep neural network as a memory module that learns what to store in real time. The model uses a surprise signal to decide what is worth keeping and what can be ignored, supported by momentum and adaptive decay to manage memory capacity. The MIRAS framework gives the theory behind this, showing how different sequence models fit into one shared design space. Early tests show strong performance on long-context benchmarks, even beating larger models on tasks that require reading millions of tokens. Google has also shown it works outside of language, including genomics and time-series tasks.

Titans hints at what comes after transformers, with models that keep learning as they run and use surprise to decide what to store or ignore. If this holds up at real scale, we may finally see AI that keeps stable context across long stretches instead of rebooting every session. Long-term memory has been the big missing piece for years, and this shift could unlock a wave of new systems that feel a lot more capable.

Transform your team’s static training materials into dynamic, engaging video guides with Guidde.

Here’s what you’ll love about Guidde:

1️⃣ Easy to Create: Turn PDFs or manuals into stunning video tutorials with a single click.
2️⃣ Easy to Update: Update video content in seconds to keep your training materials relevant.
3️⃣ Easy to Localize: Generate multilingual guides to ensure accessibility for global teams.

Empower your teammates with interactive learning.

And the best part? The browser extension is 100% free.

Check out Guidde

Evolving AI: Geoffrey Hinton thinks Google is starting to pass OpenAI, and he says the surprise is that it took this long.

Geoffrey Hinton says Google is now pulling ahead in the AI race. He points to strong reactions to Gemini 3 and Nano Banana Pro, along with Google’s push to build its own AI chips. Hinton notes that Google once led the field, inventing transformers and developing early chatbots, but held back after Microsoft’s Tay incident. He adds that Google’s scale in data centers, research teams, and upcoming chip deals with Meta strengthens its position. Hinton spoke just before Google announced a $10M CAD gift to fund a new AI chair in his name at the University of Toronto.

When Geoffrey Hinton says Google is starting to pull ahead, he is pointing at more than one strong model launch. It is about Google suddenly looking like the one player that owns the whole pipeline, from Gemini 3 and Nano Banana Pro at the top to its own TPUs that rivals like Meta may soon rent at scale. If that bet pays off, Google could end up setting the pace on model quality, prices, and safety rules while everyone else has to react.

Evolving AI: Anthropic studied how people use AI at work and found a split. AI helps many get more done, yet creatives feel pressure to hide how much they rely on it.

Anthropic tested a new tool called Anthropic Interviewer on 1,250 workers, including 125 scientists and 125 creative professionals. The tool held short conversations, and researchers analyzed the themes that came up most. Creatives say AI helps them move faster and improve output, yet many keep their use private because of how colleagues see it. Many also fear losing work to cheaper AI options. Scientists show interest in AI support but still run into reliability problems that erase the time they hoped to save. Across the broader sample, people see AI mostly as a partner, though the line between collaboration and automation is not always clear.

Underneath all these numbers is a simple story: AI is already baked into creative and knowledge work, but a lot of that use is happening in the shadows, which means the people who learn it, talk about it, and help set the rules will quietly pull ahead while everyone else is busy pretending nothing changed.

Vanguard just projected public markets may return only 5% annually over the next decade. In a 2024 report, Goldman Sachs forecasted the S&P 500 may return just 3% annually for the same time frame—stats that put current valuations in the 7th percentile of history.

Translation? The gains we’ve seen over the past few years might not continue for quite a while.

Meanwhile, another asset class—almost entirely uncorrelated to the S&P 500 historically—has overall outpaced it for decades (1995-2024), according to Masterworks data.

Masterworks lets everyday investors invest in shares of multimillion-dollar artworks by legends like Banksy, Basquiat, and Picasso.

And they’re not just buying. They’re exiting—with net annualized returns like 17.6%, 17.8%, and 21.5% among their 23 sales.*

Wall Street won’t talk about this. But the wealthy already are. Shares in new offerings can sell quickly but…

👉 My subscribers skip the waitlist

_{*Past performance is not indicative of future returns. Important Reg A disclosures: masterworks.com/cd.}

QUICK HITS

📰 Meta strikes multiple AI deals with news publishers.

🎒 Elementary school students use AI to combat homelessness.

🧬 Accurate single domain scaffolding of three nonoverlapping protein epitopes using deep learning.

🎙️ NVIDIA CEO to Joe Rogan: Nobody “really knows” AI’s endgame.

⚖️ New York Times sues AI startup for ‘illegal’ copying of millions of articles.

📱 Meta acquires AI wearables startup Limitless.

🤖 MIT researchers “speak objects into existence” using AI and robotics.

*partner link