Robotics is reaching a tipping point with In-Context Learning (ICL).
Confirming my northern star thesis of the sector turning from a hardware to compute problem. Which is where I believe the biggest growth will come from.
ICL is a new paradigm for physical AI: inference time adaptation rather than fine tuning. Tokenized demonstrations of desired behaviors (physical prompting) can be passed into the expanded context windows of the latest vision-language-action models (VLAs) and world models.
Robots can then perform complex, unique tasks in one shot, without the need for any gradient learning on the underlying model weights. World models can use ICL to simulate and plan for entirely new situations. The effectiveness of ICL depends on the quality of generalized world representations acquired during pretraining.
Meaning that rather than updating the underlying model weights for each new task, the foundation model learns the kinematic and spatial constraints of the task in its transient context state, enabling one shot learning and execution of complex tasks.
This is more in line with how LLM’s function. When you prompt a LLM chatbot asking a question, it’s not pre trained on that specific prompt with an answer for every question possible.
LLM’s are designed to use patterns and knowledge encoded within its pre trained weights alongside the information provided within its current context to generate an answer. The prompt changes what the model does without changing the underlying model itself.
Robotics has historically relied on pre trained environments to carry out a task. You can’t put a shirt in front of a robot and immediately expect it to know how to fold it. Previously, this has required 1000s of test runs in a controlled environment.
However for the first time, we’ve seen signs of ICL taking place on very short term time horizons. We saw @GeneralistAI achieve this with its Gen 1.5 - one shot learner model. The model learns with zero training on the task and a few seconds of demonstration data put into the models context. Where the robot then generalizes the task it's shown to do, called physical prompting.
The 0-1 moment everyone has been waiting for in robotics.
At the same time it's important not to jump the gun on this idea as there isn’t verifiable public evidence to confirm this, however the training models do look in line with expectations and it seems like very early signs of where the industry is heading towards.
Which if true, is where my main thesis lies for @codecopenflow and it’s SimArena.
Because if ICL and World Actions Models (WAMs) are the next frontier of scalability, it means simulations and tooling become exponentially more valuable.
Simulation will be a critical tool for Physical AI. Robotics is transitioning from a hardware limited field to a compute limited discipline. Simulators are important for both generating synthetic training data and for safely trying out new policies.
For a generalized Physical AI model to learn useful behaviors, it will need training data that covers an extremely broad range of edge cases. Collecting this data in the real world through teleoperation would take an unfeasibly long time.
Simulators solve this problem by allowing for massively parallelized data collection – you can easily run 1000s of robotic episodes at the same time in the cloud. Simulators also allow for the use of “domain randomization”, in which you add algorithmic variation to attributes like friction coefficients, lighting, mass distributions, joint tolerances etc. This is a key tool for enabling real world robot policies, trained on simulated data, aiming to solve the sim2real gap.
Codec’s SimArena biggest limitation is bounded by the fidelity of the underlying physics engines. What SimArena does is create a browser workflow environment that plugs into leading physics simulators like Nvidia Isaac Sim, MuJoCo and Genesis to test and evaluate robotic/AI foundation models.
As In Context Learning improves, SimArena can run simulations which are incredibly more accurate as robots are able to complete tasks via physical prompting (like LLM’s).
By slowly turning robotics into a compute problem, you can essentially brute force simulations that are far more accurate for real world deployment.
@Machinedotfun is the key unlock for this entire flywheel. While the launchpad was designed to enable decentralized fund raising for robotics projects and a high value network for serious builders, it’s also the bootstrapping mechanism for SimArena.
As more projects launch, provide evidence via SimArena integration and run simulations, Codec is getting first hand data to iterate and improve the workflow engine. Each machine that goes through SimArena is an additional data set that can be cross benchmarked via integrated foundation models connected to the platform. Every integration improves feedback for the workflow engine -> the engine becomes more efficient -> robots generate more volume/revenue.
We’ve already seen the first instance of this flywheel with projects like @LamportRobotics coming through Machinefun and proving evidence for their build with SimArena.
As the ecosystem expands, volume increases, revenue from the platform goes directly into buybacks for the native token. If you want my opinion, the native token being 3 mil mcap is stupidly cheap.
$CODEC coded.
Really proud of what the @Machinedotfun team has achieved in these market conditions.
Not specifically for the metrics but how we’ve approached mechanism design and the philosophy towards building.
For over a year now, major launchpads have built with the ambition for max extraction, no ecosystem support and creating PvP trading environments.
I genuinely believe what Machinefun is attempting to do is the first design that genuinely has everyones best interest in mind when you look at the launchpad holistically.
I’ll save the full mechanism breakdown for a future post but a few key variables we thought about when building are:
- Less focus on bonding curves. Lower mcap graduations, more Sol required to fill (more liquidity), less upside to be captured from launch to migration (pump is 30x vs Machinefuns 9x). Meaning there’s more incentive for traders to play at higher mcaps instead of PvP’ing sub 100k mcap coins to maximise returns.
- If there’s a tax on coins, there needs to be a positive environment created for builders, traders and the platform itself. Too many teams believe they can get away without reinvesting those rewards back into the eco and more importantly making it sustainable for the longer term (liquidity depth and trading incentives).
- Reduce token spam - higher token deployment fees & single use tickers only. Why it’s taken this long for platforms to implement this is unknown to me.
There’s quite a few features I can’t mention yet, although just these points alone would contribute to about 90% of the problems with onchain and launchpads. Most know what they are except they don’t try to fix it because traders have shown they’re willing to take sub optimal environments if there’s any sort of volume.
The great part is, this is the worst Machinefun will ever be. Everyday is focused on iterating from community feedback, developer feedback, trader feedback, project onboarding feedback etc.
We launched in the worst conditions, sentiment and volume onchain has seen in over 3 months. The idea was to launch, get something live which only gets better overtime with inputs from every vertical of interaction. Iterate, add new features and constantly improve every time we find mechanisms that align with the philosophy of creating successful long term utility runners.
Machinefun was even fortunate enough to launch on the day of the $6 Billion Binance hack and largest heist in history. Still, it hasn’t deterred the ecosystem from moving forward.
Three projects have already launched and are building in public:
1: $LAMPORT - A six legged robot (Princer1) that keeps your keys in a secure chip and signs only when you are in the room answering your hand is on it.
2: $SENSE - Building a refinement layer for robotics, allowing for open source data collection via the Lumina app.
3: $ORBIE - An open source home robot. Can complete tasks like checking on pets, the elderly or whatever you program it for.
Only been two weeks and the platform already managed to generate 70k in revenue with 25k in funding for teams who are building real products.
As the platform usability and features only become better, I imagine many more teams will be launching in the near future as this will be the one stop hub for anyone who wants to work with serious robotics, physical hardware or just general utility projects and builders.
Can’t wait for what the next few weeks have in store. If you’re a team that’s interested in launching something that aligns with the platform, my dm’s are always open for any questions.