JEPA learns what to keep from what it sees, and uses that to predict what comes next.

But what about the details it can’t predict? Wouldn’t it be easier for the model to drop them?

Look at this example
Imagine a traffic light that changes color randomly. Its next color is current color helps predict whether a cyclist will stop or move. If we drop the light, the cyclist becomes harder to predict.

▪️ A recent research call this set-level view "predictive closure:" A feature’s value depends on its relationships with the other features the model represents.

Keeping the light makes the cyclist easier to predict, but adds an unpredictable color to the model’s targets. Dropping it removes that target, but loses a useful clue.

The researchers separate these two costs:

1. Aleatoric uncertainty: what remains uncertain even with the full current observation, like the light’s next random color.
2. Predictive leakage: useful information available in the observation but missing from the representation, like the current color if the model drops the light.

An unpredictable detail can therefore be worth keeping because of what it helps predict. Random noise that helps predict nothing else offers no such benefit.

In controlled experiments, unpredictable features became easier to recover from the representation as they became more useful for predicting others.

That doesn’t guarantee JEPA keeps everything important, but it explains why unpredictable doesn’t automatically mean disposable.

It seems that DeepSeek is having its Linux moment

As Linux grew, companies like IBM helped develop it while making money from the products and services around it.

My husband, @wschenk, kept bringing up Linux, and the more I looked into DeepSeek, the more DeepSeek-Linux
Today @deepseek_ai builds DeepInfer (the software behind its inference service) on vLLM, and vLLM uses DeepSeek’s code in return.
Its open 3FS handles training data and KV caches, and developers can run and adapt DeepSeek models like V4.1 Flash themselves.
Then, moves like tuning its software for Huawei’s Ascend chips help both sides: Huawei gets more useful hardware, and DeepSeek gets another platform for models.

But why would a company share so much of the infrastructure that helps it compete? And where does competition move?

In this episode, I follow those connections to understand what DeepSeek is trying to achieve youtube.com/watch?v=SU9UW1…

There’s more to System One decision models than Jev.

These 15 models are worth a closer look:

▪️ Jeva
▪️ Prism-1
▪️ Bongard
▪️ Verdict (multi-question)
▪️ VideoJev
▪️ MJev-Spatial
▪️ AudioJev
▪️ MetaEncoder
▪️ FastJEV
▪️ SanSi (looped)
▪️ StartLux-Decision
▪️ System X d)
▪️ Intern-Decision (multimodal)
▪️ FunDM (distilled)
▪️ LAVOIR

Check out our quick guide to what each brings, how it works, and where it fits: turingpost.com/p/exploring-sy…

A dedicated resource for technical AI breakdowns and analysis – @TheTuringPost

We focus on the underlying engineering:
- guides to emerging AI systems
- research analysis
- what is happening right now

Read by 110,000+ across top AI labs, VC funds & universities

Come join us!

There’s more to System One decision models than Jev.

These 15 models are worth a closer look:

▪️ Jeva
▪️ Prism-1
▪️ Bongard
▪️ Verdict (multi-question)
▪️ VideoJev
▪️ MJev-Spatial
▪️ AudioJev
▪️ MetaEncoder
▪️ FastJEV
▪️ SanSi (looped)
▪️ StartLux-Decision
▪️ System X d)
▪️ Intern-Decision (multimodal)
▪️ FunDM (distilled)
▪️ LAVOIR

Check out our quick guide to what each brings, how it works, and where it fits: turingpost.com/p/exploring-sy…

Superintelligence in what, exactly?

Maybe it’s time to be more specific about the domains where we expect AI to go beyond human capabilities.

@MicrosoftAI makes medicine an interesting place to start with its new MAI-DxO research and medical intelligence framework↓

The study, xplores sequential diagnosis from a partial clinical picture.

MAI-DxO (Microsoft AI Diagnostic Orchestrator) uses a language model to simulate 5 clinical roles: proposing diagnoses, selecting questions and tests, challenging assumptions, checking reasoning, and weighing costs. Its assessment evolves as evidence arrives.

It achieved 83.2% accuracy on complex NEJM cases and 80.3% on emergency-department cases, vs. 36.1% and 46.8% for participating generalist physicians.

→ But what would make this medical superintelligence?

▪️ Microsoft’s framework outlines 6 dimensions of medical intelligence.

Its bar for superintelligence is high: outperforming the strongest human-led health systems in reasoning and coordination across the full course of care.

1. Medical knowledge
It applies current evidence to individual cases and recognizes what it doesn’t know.

2. Diagnostic reasoning and triage
It chooses questions, tests, and urgency, revising its view as evidence arrives. Sequential diagnosis fits here.

3. Therapeutic decision-making
It weighs benefits, risks, costs, and availability, adapting treatment as the patient responds.

4. Longitudinal context
It tracks health over time, preserving sources and uncertainty rather than starting fresh at every visit.

5. Clinical coordination
It brings in expertise, determines who acts next, and checks that follow-up happened.

6. Human values and agency
It respects patients’ goals, preferences, and right to refuse a recommendation.

These capabilities must work together. Treatment responses can change diagnoses; patient priorities can change treatment choices, etc.

That’s why a domain-specific view of superintelligence is useful - we can give a system clear tasks and criteria it needs to meet.

Harness is dead, long live the harness?

I went to @aiDotEngineer NYC and heard some very strong opinions about harnesses. Some engineers are removing parts of theirs because the models no longer need them. Others are building systems that let an agent work for days.

Of course, as to sort things out and understand what is happening with harness (and if it’s really dead).

So many questions:
- What is an agent actually doing during those days?
- How does it remember what happened without turning an earlier guess into a fact?
- And if it can build its own tools, what are we letting it change?

This episode starts with a coding agent in about 25 lines of code, and then I drew insights from @_philschmid, @dillonrolnick, @aparnadhinak, @akshat_b, and a lot from @jediahkatz who had an amazing talk.

It came out as a full guide with everything you need to know about the current state of harness

Watch the full guide on YouTube: youtube.com/watch?v=pEHfZU…

A free book: "Introduction to Machine Learning" by Laurent Younes

It’s a great read to learn math behind core ML concepts or refresh your knowledge of it, and explore some key ML methods along the way.

Covers:

- neural networks, graphical and generative models
- optimization ression, SVMs, and kernels
- decision trees
- clustering and generalization bounds

A dedicated resource for technical AI breakdowns and analysis – @TheTuringPost

We focus on the underlying engineering:
- guides to emerging AI systems
- research analysis
- what is happening right now

Read by 110,000+ across top AI labs, VC funds & universities

Come join us!

JEPA learns what to keep from what it sees, and uses that to predict what comes next.

But what about the details it can’t predict? Wouldn’t it be easier for the model to drop them?

Look at this example
Imagine a traffic light that changes color randomly. Its next color is current color helps predict whether a cyclist will stop or move. If we drop the light, the cyclist becomes harder to predict.

▪️ A recent research call this set-level view "predictive closure:" A feature’s value depends on its relationships with the other features the model represents.

Keeping the light makes the cyclist easier to predict, but adds an unpredictable color to the model’s targets. Dropping it removes that target, but loses a useful clue.

The researchers separate these two costs:

1. Aleatoric uncertainty: what remains uncertain even with the full current observation, like the light’s next random color.
2. Predictive leakage: useful information available in the observation but missing from the representation, like the current color if the model drops the light.

An unpredictable detail can therefore be worth keeping because of what it helps predict. Random noise that helps predict nothing else offers no such benefit.

In controlled experiments, unpredictable features became easier to recover from the representation as they became more useful for predicting others.

That doesn’t guarantee JEPA keeps everything important, but it explains why unpredictable doesn’t automatically mean disposable.

Looped Transformers are hot right now. 10 new papers worth your time:

- Recurrent Looped Transformer
- Looped Prediction Transformer
- Recursive Multi-Agent Systems
- Loop Flow Transformers
- and 6 more

Follow @TheTuringPost for weekly AI paper picks

AI research moves fast. We track the papers worth reading and explain the ideas behind them.

This week: Cortex, Librarian Agents, OpenGCUA, DuoDiscover, and WAVE-0.

Follow Turing Post for your weekly AI research briefing.

The interesting thing is that this super realistic voice model was trained on only one NVIDIA GB200 NVL72

Very impressive @blocks

The interesting thing is that this super realistic voice model was trained on only one NVIDIA GB200 NVL72

Very impressive @blocks