Aimee Su

Angel Investor & Founder-Operator | MIT Alumna · MIT Sandbox Innovation Fund | Co-founded FLOWERPLUS — 15M+ subscribers nationwide

[Aimee Su @AimeeSpringsong]

I’ve been reading the new Dream-RSI paper, and what I found remarkably creative is how Dream-RSI rethinks the value of past exploration: turning what looks like a sunk exploration cost into something reusable.

During online exploration, an agent searches the real environment. Its attempts, branches, evaluations, and outcomes accumulate into a Discovery Tree — History₁.

Instead of treating that history simply as a record, Dream-RSI turns completed discovery histories into replay simulators.

Then comes the clever part: offline “dreaming.��

New exploration policies can be tested by replaying History₁ rather than repeatedly paying for new online exploration. In a sense, the system “dreams” inside its own exploration history — experimenting with alternative ways of searching before returning to the real environment.

But History₁ only contains paths that have already been explored. It cannot tell the agent what lies beyond them.

So the improved Policy₂ goes back online, explores differently, potentially discovers new branches, and generates History₂. That richer history can then support another round of offline dreaming.

The recursive loop becomes:

Policy₁ → Online Exploration → History₁ → Offline Dreaming → Policy₂ → Online Exploration → History₂ → Offline Dreaming → Policy₃ → …

The underlying model can remain unchanged. What improves recursively is how the system explores, while each round of costly online exploration creates more reusable history for the next.

At a very different layer of the AI stack, this reminded me of the economic logic behind specialized infrastructure such as TPUs:

• TPUs: specialization — redesigning the computation path around the workload.
• Dream-RSI: reuse and amortization — redesigning the exploration path so previously costly search can contribute to future improvement.

#ArtificialIntelligence #AIResearch #AIAgents #DeepMind #AIEconomics

[Aimee Su @AimeeSpringsong]

@TritiumAc Very interesting direction. Particularly curious to see how these gains scale to more complex, real-world tasks—and whether this approach can meaningfully reduce the cost and time of adapting generalist policies to new skills. Looking forward to seeing how PPS develops.

[Aimee Su @AimeeSpringsong]

There’s something oddly healing about watching a robot learn — maybe because learning always looks a little clumsy at first. 😂

#Robotics #EmbodiedAI #AI

[Aimee Su @AimeeSpringsong]

@cdngdev This is so cute — and such a fun glimpse of where robotics is heading 😂

[Aimee Su @AimeeSpringsong]

A new paper from the Anthropic Institute offers an interesting way to think about a question I’ve been following closely: how does AI capability translate into economic value—and who ultimately captures it?

The paper models three economic scenarios for transformative AI through 2030:

Modest / Substantial / Extreme

GDP vs. no-AI baseline: +1.6% / +8.3% / +32.4%
Annual GDP growth: 2.4% / 5.4% / 15.4%
Labor share: 59.4% / 56.1% / 45.2%
Capital share: 40.6% / 43.9% / 54.8%
Cognitive employment: −0.5% / −3.9% / −21.5%
Cognitive unemployment: 2.9% / 4.5% / 17.9%

These are scenarios, not forecasts. The results highlight a potential divergence between aggregate economic growth, cognitive employment, and the distribution of income between labor and capital.

The mechanism behind the numbers may be even more important. Capability alone does not determine economic impact. Adoption and diffusion, automation versus augmentation, worker reallocation, and the creation of new human tasks all shape the outcome.

AI capability → adoption & diffusion → productivity → value creation → value capture

This leaves an important question: if AI creates substantial productivity gains, who ultimately captures that value? And within capital, how might it be distributed across compute, models, applications, distribution, and workflow ownership?

#ArtificialIntelligence #AIEconomy

[Aimee Su @AimeeSpringsong]

4/4

AI: defining the right problems—organizing through uncertainty and bringing human and machine intelligence together, while weighing technical possibility against economic value, social consequences, and ethical boundaries.

Full essay:
https://t.co/fheLdJClmx

#AI #Entrepreneurship #VentureCapital

[Aimee Su @AimeeSpringsong]

1/4

What kind of founder does each technological era reward?

A simple way to frame the progression—from the PC internet to mobile, and now to AI:

[Aimee Su @AimeeSpringsong]

3/4

Mobile internet: capturing emerging behaviors—moving attention, trust, and commerce onto an always-on, deeply personal device, while organizing speed, growth, and scale around a narrow window of opportunity.

[Aimee Su @AimeeSpringsong]

2/4

PC internet: seeing structure—imagining what could be digitized, then building the systems, institutions, and trust required to move activity from offline to online.

[Aimee Su @AimeeSpringsong]

The ideas explored here draw on a rapidly evolving body of work across diffusion, world models, robotics, and physical AI—including research emerging from @GoogleDeepMind, @NVIDIAAI, @StanfordAILab, and @berkeley_ai.

[Aimee Su @AimeeSpringsong]

1/ I’ve always been fascinated by technologies that begin in one domain, then quietly reveal a much larger possibility.

Diffusion models began by teaching machines to create images. But what if their deeper significance lies in helping intelligence imagine—and choose among—possible futures?

[Aimee Su @AimeeSpringsong]

4/ This is why I increasingly see diffusion as more than a media-generation technique.

It may become an architecture for reasoning under uncertainty—moving from noise, to possibilities, to choice.

My latest essay:
https://t.co/MCXAnRUwEd

#DiffusionModels #WorldModels #EmbodiedAI #PhysicalAI

[Aimee Su @AimeeSpringsong]

3/ For robotics and embodied AI, the goal may not be to predict one perfect future.

It may be to represent several plausible futures, understand how actions could change them, and identify which path is most likely to succeed.

[Aimee Su @AimeeSpringsong]

2/ The evolution is already taking shape:

• from generating pixels
• to generating video
• to modeling how environments may evolve
• to helping intelligent systems decide what to do next

Diffusion is moving from visual generation toward predictive and physical intelligence.

[Aimee Su @AimeeSpringsong]

@MartinBeraja @UCBerkeley Thanks for letting me know! I’m currently based in Boston, but I hope I’ll have an opportunity to attend in person in the future. Hope the class goes wonderfully!

[Aimee Su @AimeeSpringsong]

Really interesting framing. The performance gap seems increasingly small relative to the cost gap — which makes me wonder whether the more important question is not can LLMs replace embedding models, but where is the additional intelligence actually worth paying for? Curious whether you see this pushing toward a more explicit routing architecture: embeddings by default, LLMs only when the task complexity justifies the marginal cost.

[Aimee Su @AimeeSpringsong]

As an MIT Sloan alumna, I’m delighted to remain actively engaged in MIT’s entrepreneurial ecosystem by joining the MIT Sandbox Innovation Fund Program as a mentor.

Looking forward to supporting MIT founders and scientists in moving frontier technologies from idea to venture to impact—connecting technology, people, business, and society. ����

#MITSandbox #Innovation #Entrepreneurship #TechnologyCommercialization

[Aimee Su @AimeeSpringsong]

@PeterWBattaglia @GoogleDeepMind @GoogleResearch The color palette immediately caught my eye—beautiful, intuitive, and remarkably effective at making complex weather data feel accessible.