Abinav Ramesh

i talk to machines for a living
physical AI, shop-floor data, pharma machines
100 days of physical AI companies

[Abinav Ramesh @abinavmoto]

100 days of physical AI companies, daily

what i've learned connecting machines to data for 8-10 years in India:

> many teams find out what's broken only after it stops
> the operator's notebook is the most honest software review
> AI on data nobody trusts just makes bad decisions faster

here i post
> a physical AI company every day
> Industrial IOT

[Abinav Ramesh @abinavmoto]

Extra mile achievements that a startup can do ,,…

[Rajan Anandan @RajanAnandan]

For decades, India’s manufacturing story was about low-cost production and supply-chain coordination.

A new cohort of startups is moving past that, competing in high-precision, IP-led manufacturing and building complex products for customers around the world.

Jeh Aerospace is one of them. In the 18 months after its seed round, the startup secured $200 million in long-term contracts from global customers!

Precision manufacturing is a Trillion dollar opportunity that India has barely tapped. @jehaero built by @vishalsanghavi and Venky Mudragalla is an early sign of what’s coming.

Read more about this company and many more incredible startups in The Lift-Off Nation, co-authored by @TheSaurabhSri and me.

[Abinav Ramesh @abinavmoto]

asim's pitch is find the failure before the robot meets people

not one glass pick that worked on camera
they say they vary shape, weight, friction, material on real robots

What is climed claimed:

standardized physical-intelligence benchmarks proven in public so far: launch video + YC copy

[Tristan Lecourtois @tristanlecourt]

We’re launching Asim (YC F26) to help robotics teams find failures before deployment.

A robot completes a task perfectly. Can it handle an unfamiliar object, reason about how parts fit together, or recover when an interaction goes wrong?

We’re building standardized benchmarks for physical intelligence. We test real robots through repeatable scenarios that probe the capabilities behind successful actions: spatial reasoning, understanding physical interactions, dexterous manipulation and adaptation.

Teams can identify where performance breaks down, get actionable engineering feedback and measure whether the next software update actually improves it.

Building robots? Let’s put them to the test → https://t.co/yVIWlWoYln

Asim. Benchmarking physical intelligence.

[Abinav Ramesh @abinavmoto]

They are pulling off features everyday …

[Perplexity @perplexity_ai]

Introducing teammates.

Teammates work for you. They take on jobs, and work proactively using your Computer tasks, skills, memory, and tools. Teammates collaborate with other teammates and can message coworkers in Slack.

Available now for Max subscribers on the web.

[Abinav Ramesh @abinavmoto]

@kimmonismus If people keep saying no to data centers, smaller models that run closer to the user suddenly look like the real value for money !!

[Abinav Ramesh @abinavmoto]

@nio_robotics No body just arms, bold !!
Would love to see it climbing a tower, not just standing in the field

[Abinav Ramesh @abinavmoto]

@thejerrycheng Open source ecosystem is getting better and bigger for the Physical AI growth

[Abinav Ramesh @abinavmoto]

day 17/100 of physical ai companies

most robots learn a task by being shown it on a robot, many times

@liberai_robot from beijing uses a cheaper teacher.. that is “people”

humans wear a thin data glove, cameras record the work, and one model turns it into robot moves

their own test: 60 tasks it never practiced, 600 tries, 36% success

pick and place 63.5%, multi-step chores 10.5%

if human data can stand in for robot demos, teaching robots gets far cheaper

the 10.5% is the number to watch…

#100DaysOfPhysicalAI #MachinesThatThink

[LiberAI @liberai_robot]

Introducing Liber-0, our World Action Model for zero-shot robot manipulation—including dexterous, multi-step tasks.

Across 60 tasks proposed by GPT-6 and 600 trials, Liber-0 achieved 36% overall success without robot demonstrations collected specifically for those tasks:
• Pick-and-place: 63.5%
• Other manipulation tasks: 34.0%
• Challenging multi-step tasks: 10.5%

Zero-shot capability extends beyond individual actions to coordinated sequences.

Pretrained on 2.2M+ hours of human experience, Liber-0 jointly predicts future observations and actions to turn human experience into robot skills.

Technical blog: https://t.co/X5NMTjsr8r

[Abinav Ramesh @abinavmoto]

@Nakul_Agarwal04 @NeuralMiles My take is like there could be operator shortage , harsh environments that operators cannot be at vehicles kind of things that got into my mind

Will it be possible like robots if we are able to teach sequence and outcome it completes the task or job ?

[Abinav Ramesh @abinavmoto]

We spent years making AI understand what we say.

But understanding words and having a conversation are two different things.

Think about it.

When we talk to someone, we don’t just listen to their words. We notice their expressions, when they hesitate, how they react and when they interrupt.

Most AI today still misses that part.

That’s what caught my attention about @Atmanity

Its founder, @yuchenping previously built a startup that was acquired by Google. Now he’s back with something interesting.

They just launched Atmee v2 @Atmanity

The idea is simple. Give it an image and a short audio sample, and it can generate an AI character that interacts with you in real time.

Not just speaking.

Listening, reacting, showing expressions and responding when you interrupt.

What makes this interesting is the engineering behind it.

Instead of making one large model handle everything, they’ve separated conversational behaviour from visual generation.

One model decides how the character should behave, while another generates the expressions and movements.

They claim this approach is up to 50x faster and 10–100x cheaper than general-purpose video models.

Those numbers still need validation at scale.

But I think the bigger story is not about AI avatars.

It’s about the interface.

Today we open a chat window, type something and wait for AI to respond.

Tomorrow, interacting with AI might feel more like talking to another person.

Imagine a tutor who notices you’re confused, a training assistant that reacts to your mistakes or a digital operator who can guide someone through a complex process.

There are still questions around latency, realism, trust and how useful these characters will be beyond demos.

But one thing is becoming clear.

We are moving from AI that can answer us to AI that can interact with us.

And that changes how we think about building products around AI.

Worth watching what Atmanity builds next.

[Chen-Ping Yu @yuchenping]

After Google acquired my AR startup, I started thinking about the next interface, this time for AI. Text and voice work fine, just like DOS did before the Mac. But nothing’s more intuitive than face to face.

After two years of research, today we’re launching Atmee v2: a real-time avatar model that makes live video agents feel truly life-like.

• Any character from 1 image + 10s of audio
• Contextual expressions, gestures, even interacting with the environment
• Live video calls, voice only, or video generation

At its core, Atmee v2 is a video foundation model, it can render anything, but it’s built to excel at characters, which is what video models like Seedance 2.5 and Wan 3.0 are mostly used for. In comparison:
Atmee v2

• Up to 50x faster
• 10-100x cheaper

Build tutors, sales coaches, support agents, companions, game characters, even whole living worlds. The interface after text and voice starts today.

[Abinav Ramesh @abinavmoto]

@abulhasanx @neosigma_ai @gauri__gupta Exactly , one liner for product!!

[Abinav Ramesh @abinavmoto]

An AI agent completed a task.

Or at least, that’s what it told the customer.

The file it created was wrong.

And that’s exactly the problem one AI startup is trying to solve.

Meet @neosigma_ai founded by @gauri__gupta , a San Francisco startup building AI agents that improve from their own mistakes.

Here’s what caught my attention.

- They studied 928 real production sessions from one customer’s AI agent.

- 80 sessions failed to deliver what the user wanted.

- 67 were mistakes made by the agent itself.

Some generated broken file links, Some returned templates instead of completed reports.

Others confidently made claims the data didn’t support.

And the interesting part?

@neosigma_ai turned those failures into 16 reusable evaluation tests.

Those tests covered 42 out of 51 previously observed agent failures.

Now let’s Think about that….

We are spending billions making AI models smarter.

But what happens after these models enter the real world?

- A customer reports a problem.
- An engineer investigates.
- Someone changes a prompt.
- A new version goes live.

And the same mistake can happen again.

@neosigma_ai is trying to change that loop.

→ Observe real agent failures.
→ Understand what went wrong.
→ Recreate the failure in a test environment.
→ Validate the next version before deployment.
→ Keep learning from new failures.

They also published an experiment where agent task success went from 56% to 78%, without upgrading the underlying model.

Of course, benchmark improvement isn’t the same as proven reliability at enterprise scale.

But the direction is interesting is what I see

The next big advantage in AI may not come from having the smartest model.

It may come from having a system that remembers what went wrong and makes sure the next version doesn’t repeat it as I come from a industrial environment where data is abundant

Every company deploying AI agents is generating something valuable

Not just data. It’s Experience.

And the companies that can turn that experience into measurable improvement will have an advantage that’s difficult to copy in this fast paced market.

#AI #AIAgents #Startups #AgenticAI

[Gauri Gupta @gauri__gupta]

The intelligence frontier keeps advancing every day. Are your agents delivering better outcomes on what matters?

For enterprises, the promise of agents is more work completed reliably, with less human intervention and a lower cost per successful task.

Today, we’re launching @neosigma_ai and our mission to empower every organization to build their own sovereign intelligence lab.

A place to operationalize their knowledge, replay real work, and optimize agents against the outcomes that matter to their business. Where production experience compounds into intelligence the organization owns.

We’re starting with NeoEval, our automated evaluation engine. It turns your production data into evaluation environments - that dynamically evolves along with your users, workflows, and frontier.

A playground for evaluating and optimizing your agents.

Your data. Your institutional knowledge. Now operationalized into evals and compounding intelligence.

Stay tuned for the next few days of updates from NeoSigma.

[Abinav Ramesh @abinavmoto]

@AravSrinivas The 51% → 86% jump is interesting, and that helps in lot of ways

How this is making those decisions verifiable, not just more accurate on a benchmark ?

[Abinav Ramesh @abinavmoto]

@perplexity_ai Exact real usecases we are talking about ..

[Abinav Ramesh @abinavmoto]

@yuchenping This helps in creating many interesting content that can be used in YouTube tutorial videos

[Abinav Ramesh @abinavmoto]

the first thing that breaks when you connect machines to a dashboard isn't the network

it's tag names

> one machine says TT101
> the next says Temp_1
> the third says R1_PV_TEMP

same temperature, three names, and a year later nobody remembers which is which

fix it before the first byte leaves the PLC

> one naming rule for every site
> units in the tag description
> a tag list operators can actually read

boring work.. saves a lot of pain later

[Abinav Ramesh @abinavmoto]

@Bhavani_00007 @Meituan_LongCat Chinese products compete with rest of the world in Quality

Which Chinese products were banned !!

In AI run they are hitting hard in both hardware and software markets

[Abinav Ramesh @abinavmoto]

@AskPerplexity Playing with DIY stuff is fun but making it to real life usecases that might save 1000’s of dollars will be a get value for money!

[Abinav Ramesh @abinavmoto]

Capital can move globally. Trust and context don’t move that easily.

For investors, local presence isn’t just about sourcing deals. It’s about building proprietary deal flow, developing conviction before consensus forms, and winning allocation in competitive rounds.

You can invest remotely. But can you consistently spot and win the best outlier founders before everyone else does?

If local presence is no longer essential, where does the next generation of proprietary deal flow and investor alpha come from?

Curious how VCs are thinking about this: Is physical proximity becoming obsolete, or are we confusing access to deals with an actual investing edge?

[Abinav Ramesh @abinavmoto]

25,000 humanoids shipped in the first half of 2026.. more than all of 2025

but here's the part nobody quotes

about 70% went to labs, schools, exhibitions and data-collection centres

only ~30% went into real commercial work

and GXO, one of the biggest users, says it's still a couple of years from ROI

robots have left the lab.. most of them just haven't reached the floor yet

i broke down the full numbers here 👇

#MachinesThatThink

[Abinav Ramesh @abinavmoto]

https://t.co/6883eSaL0Y