The standard makes sense: models work much better when devices expose standardized, machine-readable interfaces.
But as far as I can tell, this technical "standard" isn't actually public yet (you have to apply for access to even see or implement it.) Anthropic says they plan to open source it later.
Quite a departure from the way foundational hardware standards like USB and CAN were developed. You shouldn't need permission to read a standard.
MCP was a bucket of Not Invented Here nonsense, that ignored years of protocol design. It was incredibly jank and hard to run at any scale, for years and years.
They finally looked at how other protocols do things, and made something that behaves in recent versions. But it's hard as heck to take this company at all seriously. The don't really care about ecosystems (AGENTS.md), they don't seem to know think or care much about protocols until they have to (MCP's incredibly long redemption ark). The vibes are all off.
Maybe they avoid laying their own pitfalls this time. Maybe it's better. It'd be nice to start seeing some kind of cause to trust this company as vaguely capable computer-using people. But so far that hasn't been the case.
I think they're doing this so they can have months of advanced notice of how the protocol will look, to post-train their models to be the only ones with the ability to interact with this protocol on the day it becomes public.
It's like creating conditions for a race where they structurally have months of advance. And realistically they have a credible shot at establishing the standard due to their weight in the industry, and control of the most popular AI distribution method (Claude).
So they create the race, knowing the race is what will likely establish the standard, and using some pretext to prevent other competitive athletes from training for it in advance. Other athletes only get to start when the gun fires.
I think I agree. I find it a little bit annoying that they do not start by a survey of the existing landscape, discuss the issues and what works well, and spend the effort to show that they can indeed to be trusted as the authority behind such a protocol expected to become standard. For MCP, it started very messy, but from what I understand it was developed by a small team within the company working on internal tooling, for their own use case, I think it’s reasonable to expect something a bit awkward, I don’t think they expected its success
It's a bit of both for Anthropic I think, sometimes cutting edge and quite interesting or just good improvements, sometimes ignoring best practices either recently established or known for decades. Obvious to see where the smart people are in high places at the company.
To me, this seems like OPC UA / SiLA but instead of being software <-> machine semantics/control it's AI <-> machine semantics/control. Or in simpler terms, an AI-facing hardware abstraction / device-description standard.
MHS can describe capabilites, metadata, safety limits, as well as provide read/write control and discovery. Things like "can measure temperature", "arm weighs X kg", "never exceed X RPM", that agents can easily understand. (as opposed to that being buried in a datasheet somewhere, or having to be included in the prompt)
To actually add something to the discussion though: I guess getting LLMs to control robots is cool, but would potentially have more severe consequences than pure software since it interacts more with the physical world
There are MANY packages in this space: Labs love to invent their own versions powered by the hubris of grad students. Anyway a random sampling of some I’ve come across is:
The general problem is (1) i have a heterogeneous mix of hardware that speaks a bunch of different protocols from SCPI serial commands to VISA over USB, to bizarre proprietary things and (2) I want to be able to talk to it from one central computer and (3) ideally have some level of standardisation or abstraction so I can express an experiment like move <stage> from <x> to <y> in <n> steps and at each step fire <laser> and measure <signal> while logging <temperature>.
Anthropic seem to be creating the n+1 version of those packages and presumably doing some harness work to make Claude good at using it.
More like MCP for robotics. IMO they are better off getting the robots to use existing human UIs than trying to get hardware manufacturers to play ball.
It's hard to make LLMs read the manual for an invented language and have them improvise on the spot(1-shot). An established language or a command set that can be included in the training set to be used off top of their head is a lot more reliable and efficient.
The caveat is that actual utility of this particular command set isn't clear yet. Or how many LLM labs will find it worthwhile.
In the context of lab hardware there’s a ton of proprietary software and barriers to automating things. I think this is less about how to communicate, and more guaranteeing that “yes, this piece of hardware can interface with an agent” and is meant to do so. Kind of like using MCP vs letting your agent make raw HTTP requests
It seems more like a basic instruction set, that will then be translated to something actionable for each different device. Seems like most of the work then will be in developing the drivers for whatever devices you want the LLM to access.
So far Anthropic shows remarkable taste in research, engineering, and product. Dario the ideologist attracts the most pure-minded talent in the industry. It indeed looks like Anthropic would be unchallenged for the foreseeable future. It'll be very interesting to see its downfall in 5-20 years.
The standard makes sense: models work much better when devices expose standardized, machine-readable interfaces.
But as far as I can tell, this technical "standard" isn't actually public yet (you have to apply for access to even see or implement it.) Anthropic says they plan to open source it later.
Quite a departure from the way foundational hardware standards like USB and CAN were developed. You shouldn't need permission to read a standard.
As far as I can tell is these standards (MCP/MHS/etc) are just semi obvious tool inferfaces that Anthropic uses as training scenarios.
This contrasts nicely with all the breathless high-profile security incident write-ups:
Our model tricked us, and tried to extort users to achieve its ends! Their model hacked HuggingFace! These things aren't safe!
Anyways, here's a hardware interop protocol for AI we're hoping becomes a standard. Hey look, you can even hook up microscopes!
MCP was a bucket of Not Invented Here nonsense, that ignored years of protocol design. It was incredibly jank and hard to run at any scale, for years and years.
They finally looked at how other protocols do things, and made something that behaves in recent versions. But it's hard as heck to take this company at all seriously. The don't really care about ecosystems (AGENTS.md), they don't seem to know think or care much about protocols until they have to (MCP's incredibly long redemption ark). The vibes are all off.
Maybe they avoid laying their own pitfalls this time. Maybe it's better. It'd be nice to start seeing some kind of cause to trust this company as vaguely capable computer-using people. But so far that hasn't been the case.
I think they're doing this so they can have months of advanced notice of how the protocol will look, to post-train their models to be the only ones with the ability to interact with this protocol on the day it becomes public.
It's like creating conditions for a race where they structurally have months of advance. And realistically they have a credible shot at establishing the standard due to their weight in the industry, and control of the most popular AI distribution method (Claude).
So they create the race, knowing the race is what will likely establish the standard, and using some pretext to prevent other competitive athletes from training for it in advance. Other athletes only get to start when the gun fires.
I think I agree. I find it a little bit annoying that they do not start by a survey of the existing landscape, discuss the issues and what works well, and spend the effort to show that they can indeed to be trusted as the authority behind such a protocol expected to become standard. For MCP, it started very messy, but from what I understand it was developed by a small team within the company working on internal tooling, for their own use case, I think it’s reasonable to expect something a bit awkward, I don’t think they expected its success
It's a bit of both for Anthropic I think, sometimes cutting edge and quite interesting or just good improvements, sometimes ignoring best practices either recently established or known for decades. Obvious to see where the smart people are in high places at the company.
Sounds like PyLabRobot:
https://docs.pylabrobot.org/stable/
How does this compare to ROS 2? Roughly the same goals.
How is this different from tool calls? Or rather, MCP?
It’s actually really cool, and not just for AI. That should make it easy to automate some processes with simple scripts.
Nitpick: it’s annoying that they add Model to those acronyms, for what seems like a fairly universal software interface
To me, this seems like OPC UA / SiLA but instead of being software <-> machine semantics/control it's AI <-> machine semantics/control. Or in simpler terms, an AI-facing hardware abstraction / device-description standard.
This is how I understand it:
Agent <-> MCP/CLI/code <-> MHS <-> vendor API/SCPI/OPC UA/ROS/etc <-> CAN/Modbus/USB/etc <-> hardware
MHS can describe capabilites, metadata, safety limits, as well as provide read/write control and discovery. Things like "can measure temperature", "arm weighs X kg", "never exceed X RPM", that agents can easily understand. (as opposed to that being buried in a datasheet somewhere, or having to be included in the prompt)
Also see: https://en.wikipedia.org/wiki/OPC_Unified_Architecture, https://en.wikipedia.org/wiki/Standardization_in_Lab_Automat..., https://xkcd.com/927/
Cool story Dario, call me when Claude supports AGENTS.md.
If you can't prompt Claude to set up a symlink you probably shouldn't be using Claude.
If they can’t prompt Claude to implement such a trivial standard, they probably shouldn’t be proposing standards.
As the saying goes every 5 minutes in Silicon Valley another company reinvents ROS
https://xkcd.com/927/
To actually add something to the discussion though: I guess getting LLMs to control robots is cool, but would potentially have more severe consequences than pure software since it interacts more with the physical world
I wonder if there will be an updated xkcd for ai.
Like... don't need standards, just explain to an ai what the target hardware is and it will output binary bits that match.
Hardware-speech synthesis breakdown and decline
I don't want to be that guy but did they invent bluetooth for ai? Like what is MHS actually? I read it twice and still don't get it
There are MANY packages in this space: Labs love to invent their own versions powered by the hubris of grad students. Anyway a random sampling of some I’ve come across is:
The general problem is (1) i have a heterogeneous mix of hardware that speaks a bunch of different protocols from SCPI serial commands to VISA over USB, to bizarre proprietary things and (2) I want to be able to talk to it from one central computer and (3) ideally have some level of standardisation or abstraction so I can express an experiment like move <stage> from <x> to <y> in <n> steps and at each step fire <laser> and measure <signal> while logging <temperature>.Anthropic seem to be creating the n+1 version of those packages and presumably doing some harness work to make Claude good at using it.
More like MCP for robotics. IMO they are better off getting the robots to use existing human UIs than trying to get hardware manufacturers to play ball.
Oh I see, yeah I guess this assumes manufacturers will just adopt this standard universally. Clears things a bit!
What is the difference of MHS to MCP when it comes to features?
The whole text reads like everything is just a gRPC call that could've also been implemented with an MCP based wrapper.
That 2 trillion dollar valuation ain't gonna make itself.
It's hard to make LLMs read the manual for an invented language and have them improvise on the spot(1-shot). An established language or a command set that can be included in the training set to be used off top of their head is a lot more reliable and efficient.
The caveat is that actual utility of this particular command set isn't clear yet. Or how many LLM labs will find it worthwhile.
In the context of lab hardware there’s a ton of proprietary software and barriers to automating things. I think this is less about how to communicate, and more guaranteeing that “yes, this piece of hardware can interface with an agent” and is meant to do so. Kind of like using MCP vs letting your agent make raw HTTP requests
It seems more like a basic instruction set, that will then be translated to something actionable for each different device. Seems like most of the work then will be in developing the drivers for whatever devices you want the LLM to access.
Yes, it turns the hardware driver problem into a translation problem (translate the hardware manual/datasheet into this protocol.)
Guess what's really good/quick at translating things?
"standard" is doing a lot of work here
Are they in a position to define any kind of standards?
So far Anthropic shows remarkable taste in research, engineering, and product. Dario the ideologist attracts the most pure-minded talent in the industry. It indeed looks like Anthropic would be unchallenged for the foreseeable future. It'll be very interesting to see its downfall in 5-20 years.
5-20 years?
Have you seen the financials of the AI companies?
Do you have inside information we don't?
OAI and ANT are private, so I sure haven‘t seen their books.