kishan

AI Engineer at Spotify, NYC

[kishan @boardgamesio]

@herb_dev Dope

[kishan @boardgamesio]

@cormachayden_ @oasishealthapp What store?

[Bryan Johnson @bryan_johnson]

I just launched an AI model based on sleep data… and it accurately predicts your age.

I teamed up with @m_franceschetti, and it's now available on their platforms.

It turns out you have a sleep fingerprint. This research can identify you out of thousands of users from one night's signal with 92.5% accuracy.

It also detects…
+ biological age within 3.3 years
+ diabetes better than Apple’s model
+ speed of aging
+ heart failure at 0.822

This Eight Sleep model is possibly the most accurate contactless bioage estimate ever reported.

What we did:

#1 What data was it built on?

This model was built on the largest raw biosignal dataset ever used to train an AI, from any device, including every wearable on the market.

> 2.04 million hours
> 136,575 participants
> 498k sessions
> 122 million segments

#2 What can it predict?

It can predict your biological age (the age your body acts like) within 3.3 years.

It also predicts heart-related and metabolic conditions. Here are the detection scores (AUROC):

> diabetes (0.852)
> heart failure (0.82)
> hypertension (0.810)
> sleep apnea (0.792)
> snoring (0.751)
> general heart conditions (0.734)
> cancer (0.678)
> hot flashes (0.671)
> migraines (0.673)

#3 How was it built?

Interestingly, the pretraining task was not “predict someone’s age”. The model was tasked with comparing two 60 second windows across different nights to figure out if the nights belonged to the same person.

To do that, it had to find someone’s ‘sleep fingerprint’. These are biological signals that the data is coming from the same person. Things like how forcefully your heart contracts, your breathing depth and rhythm, and the timing of the recoil wave each heartbeat sends through your body.

Those signals are age-predictive. It learned to estimate age, detect diabetes, and flag heart failure as a downstream readout.

The whole pretraining run was ~four days.

#4 Why is it good at age?

The reason aging prediction is accurate is mechanical. Aging stiffens arteries, reduces cardiac compliance, changes autonomic tone, HRV declines. Aging also alters sleep architecture. Deep sleep shrinks and fragmentation rises.

Every one of those changes the recoil waveform and its overnight dynamics.

Said differently, the heart of a 65 year old mechanically pushes the body differently than a 25 year old's.

#5 Why a bed vs wearables?

A bed is an elegant solution. It makes a high fidelity uninterrupted 5 to 10 hour recording every single night possible. And session-level sequence modeling becomes viable.

Whereas wearables get fragmented data: battery limits, sparse snippets, people taking the watch off, adherence dropping over weeks.

#6 More data, better prediction

The bigger the training batches (the more people the model compares at once) the better it got, log-linearly (R²=0.982).

That means the recipe is predictable: you can forecast improvement with more compute, the same way scaling laws work for language models.

The current model only ever compares two nights at a time, and the average training user contributed under 4 nights.

The team's stated next step is modeling 30+ consecutive nights per person. You can imagine how this will improve with the constant stream of data Eight Sleep gets every night.

—

It’s worth noting some limitations. Internal labels are self-reported and external cohorts are small, and that this is a research milestone, not a diagnostic device.

What makes this exciting: a passive, daily activity like sleep can now provide meaningful insight into your well-being.

I speak often of Autonomous Health, a world where the things around us take care of us without our knowing or asking.

Eight Sleep is a great example of this in practice and a major reason I maintain so much optimism for the future of health.

[Janel Comeau 🍁 @VeryBadLlama]

the most important step in getting healthy is figuring out the unique combination of weird shit you're going to mix into Greek yogurt for the rest of your life

[kishan @boardgamesio]

@andyfang 😂

[kishan @boardgamesio]

@TTrimoreau Agree

[kishan @boardgamesio]

@rikulauttia They fail to understand the details that weren’t provided to them. People make agentic tasks seam like their a magic one click button, so people take shortcuts and use them in effectively. It’s hard to be patient and use them with disciplines because they feel like magic.

[kishan @boardgamesio]

@soham_nayak04 Fable better

[kishan @boardgamesio]

@amaashvi You’d have to use vibe coded software as the training set which would be cannibalism

[kishan @boardgamesio]

@ben_sage True. Black jack + counting cards.

[kishan @boardgamesio]

@mattpocockuk Love wayfinder

[kishan @boardgamesio]

@Jbm_dev You gotta use the not interested button more

[kishan @boardgamesio]

@_coenen And touch grass

[kishan @boardgamesio]

@corbin_braun Creating real impact with AI is not as intuitive as marketed

[kishan @boardgamesio]

@icanvardar Make the AI models argue about web frameworks

[kishan @boardgamesio]

@phibrowser 100%. Garbage in, garbage out. I tried to make it super human friendly. Flip thru the slides and add what you want, drop what’s noise. Saves me time!

[kishan @boardgamesio]

“send this to an agent” for Google Slides sucks.

Slides are full of noisy artifacts, images, charts, and other visual junk that don’t convert cleanly into LLM context.

Built a simple human-in-the-loop tool today for turning decks/sites into agent-ready markdown.

It extracts slide text by default, lets you deselect noisy parts, magic-wand only the images that matter, and append image summaries directly into the markdown.

Better context → better agent output.

[kishan @boardgamesio]

@jyriso I will play your hunger games sir

[kishan @boardgamesio]

@steipete Why not just Opus

[kishan @boardgamesio]

Fable orchestration + Opus execution saves me hundreds

[Ryan Hoover @rrhoover]

I don’t understand the argument that LLMs can’t come up with original ideas.

“It’s just trained on existing data, predicting the next token.”

How is that different than human ideation?

We’re also pattern-matching machines, influenced by our life experience.