Marcos

Founding & ML Engineer at @stochasticai

Efficient finetuning:

[Marcos @marcos_98_rm]

I'm really in love with xMagic Composer ♥️. I can quickly build an agent as if I were chatting with a friend, instead of going through a mess of nodes, configurations and connectors

[Stochastic @stochasticai]

AI is supposed to make things faster. So why does building with it still feel this slow?

Most teams don't stall because of the technology. They stall because of everything around it.

[Stochastic @stochasticai]

Everyone's locking down where their files are stored.

The harder part is everything an AI agent creates while it works. Ask it to prepare a prior auth packet and it reads clinical notes and insurance details nobody pasted into a chat. The prompt, the model's response, the logs, all of it carries the same sensitive info.

So the questions change:
->Where does the model actually process your records?
->Which services keep a copy of the prompts and outputs?
->How long is it retained, and can it be used for training?

With xMagic, we split control from compute. Stochastic runs the control plane, which stores only metadata and usage. The compute plane runs in your cloud, so agents, models, files, prompts and responses stay there. You can use our models in your cloud or bring your own open source one, and optional fine tuning stays in your cloud too.

Your institution's own intelligence, running inside the environment you control.

Read the full post: https://t.co/OB1LKHGRvz

Get early access: https://t.co/4CC7nUN5U3

[Stochastic @stochasticai]

Everyone's "using AI" in some form. Nearly 9 in 10 orgs, by recent numbers.

Fewer than a quarter have actually scaled an agent into production.

That gap is basically the whole story of enterprise AI right now. Healthcare is even further behind, with production agent adoption around 18% compared with roughly 47% in banking and insurance.

Getting AI to answer isn't the hard part. Getting it to execute is, and that's where the questions change:

Can it act inside a claims system, an EHR or a payer portal, not just a chat window?
Does it know when it's unsure and hand off to a person with the full context ready?
Is every action it takes logged and traceable?

We wrote about what's actually different in the orgs that closed the gap.

Your institution's own intelligence.

Read the full post: https://t.co/etr12XwwGv

Get early access: https://t.co/4CC7nUN5U3

[Stochastic @stochasticai]

Everyone's renting the same model. The edge isn't the model. It's what only your institution has: the decisions, corrections, and exceptions no one else sees.

We call this institutional intelligence, an AI workforce that's adaptive, runs inside your trust boundary, and compounds the longer it runs. Not a tool that's the same in your hands as your competitor's.

Thomson Reuters, Harvey, and Cursor all just shipped their own models built this way. We built xMagic the same way, for healthcare.

Full piece: https://t.co/XW9yyZAhYQ

Be among the first to experience xMagic. Get early access: https://t.co/4CC7nUN5U3

Your institution's own intelligence.
-The Stochastic Team

[Stochastic @stochasticai]

Voice agents ace demos, then break in production: they hallucinate appointment slots from messy prompts, keep running on stale assumptions after a user correction, and lose track of your first request when you ask a second.

We looked at 3 failure modes that show up again and again:
→ Hallucinated data when agents rely on unstructured prompts
→ Stale decisions after a user corrects themselves
→ Dropped tasks when users multitask

xMagic's talker-and-reasoner architecture fixes all three: deterministic data retrieval, live state recomputation, parallel task handling.

Full breakdown: https://t.co/cZGzay3B4U

Be among the first to experience xMagic. Get early access: https://t.co/4CC7nUMy4v

[Stochastic @stochasticai]

Everyone's connecting AI agents to tools with MCP.

Connecting isn't the hard part. Operating that agent safely once it touches real business systems is.

Once agents go live, the questions change:

Can this agent approve a refund?
Who approved a given action, and can we audit it?
If something goes wrong, how fast can we recover?

MCP servers give agents access to tools and data. Skills define how they use that access consistently. Neither one makes an agent production-ready. The missing piece: guardrails, human approval for sensitive actions, isolated execution, observability, and controlled versioning.

Read the full blog: https://t.co/5bjB5WoYru

Be among the first to experience xMagic. Get early access: https://t.co/4CC7nUN5U3

[Stochastic @stochasticai]

What does it actually look like to build an AI agent in minutes?

We built xMagic's Composer to answer exactly that. Describe what you want, watch it take shape, test it instantly — no configuration, no switching tools, no infrastructure to set up first.

From idea to production-ready AI agent, in one connected workflow.

Get Early Access - https://t.co/u2jJvClSHV

[Marcos @marcos_98_rm]

Here's a quick demo of what building an AI agent with xMagic Composer looks like

[Stochastic @stochasticai]

What does it actually look like to build an AI agent in minutes?

We built xMagic's Composer to answer exactly that. Describe what you want, watch it take shape, test it instantly — no configuration, no switching tools, no infrastructure to set up first.

From idea to production-ready AI agent, in one connected workflow.

Get Early Access - https://t.co/u2jJvClSHV

[Stochastic @stochasticai]

The first version of an AI agent is rarely the one that makes it into production.

And that is completely normal.

The real challenge is not getting an agent to work once. It is being able to test an idea, understand what went wrong, make a change and validate the result without turning every iteration into a small engineering project.

That feedback loop is still far too slow for most teams.

With Composer, you can describe a change in natural language, test the updated behavior immediately and keep refining the agent in the same place. No jumping between environments. No redeploying just to validate a small adjustment. No manually reconstructing previous versions when something breaks.

Building reliable AI agents is not about designing everything perfectly upfront.

It is about making iterations fast enough that you can continuously improve them.

We wrote more about how we are approaching this at Stochastic:
https://t.co/Z9Gcm9HYYd

Are you interested in xMagic? Get early access here:
https://t.co/4CC7nUMy4v

[Marcos @marcos_98_rm]

The best AI agents are not built in a single shot. They improve through fast iterations: describe the change in natural language, test the updated behavior right away and keep refining from there

That is the workflow we are building with xMagic Composer

[Stochastic @stochasticai]

The first version of an AI agent is rarely the one that makes it into production.

And that is completely normal.

The real challenge is not getting an agent to work once. It is being able to test an idea, understand what went wrong, make a change and validate the result without turning every iteration into a small engineering project.

That feedback loop is still far too slow for most teams.

With Composer, you can describe a change in natural language, test the updated behavior immediately and keep refining the agent in the same place. No jumping between environments. No redeploying just to validate a small adjustment. No manually reconstructing previous versions when something breaks.

Building reliable AI agents is not about designing everything perfectly upfront.

It is about making iterations fast enough that you can continuously improve them.

We wrote more about how we are approaching this at Stochastic:
https://t.co/Z9Gcm9HYYd

Are you interested in xMagic? Get early access here:
https://t.co/4CC7nUMy4v

[Sarthak Langde @sarthaklangde]

🔥 Exclusive access for beta testing! 🎉 Be among the privileged few to try out @stochasticai xFinance, the model that outperformed BloombergGPT.

🔐 Secure your spot: https://t.co/MeE07uTZug

#LLM #finance #xTuring

[Stochastic @stochasticai]

To read more, click here: https://t.co/i8ftWcque6

#ai #ChatGPT #Bloomberg #opensource #LLMs

[Stochastic @stochasticai]

Recently, Stochastic conducted an experiment to see if it could outperform a BloombergGPT model, using a model 4x smaller, $1000, and publicly available data.

The results were shocking.

[Sarthak Langde @sarthaklangde]

Check out @stochasticai xFinance model that outperforms 4x larger BloombergGPT on finance tasks. All for less than $1000.

Blog: https://t.co/bzSc2wy5nB

https://t.co/QlqP0iGGQM

[Stochastic @stochasticai]

Learn how to easily fine tune GPT like models all on your personal laptop.

The days of needing super expensive hardware for building your own AI are gone. xTuring easily allows INT4 fine-tuning of LLMs with only 6GB of memory.

We implemented the open source tool... (1/3)

[Stochastic @stochasticai]

Exciting updates xTuring 0.0.7🙌

- Model hub for fine-tuned models
- FastAPI server for model deployment
- Dataset generation from directory of files
- Text generation using Cohere and AI 21
- Minor fixes with model loading

Great start thanks to you all- we have 800+ stars

[Glenn Ko @glennko]

Thank you @Harvard @innovationlab for selecting @stochasticai as one of the finalists for the 2023 President's Innovation Challenge!

Come see me present on 5/3.

[Glenn Ko @glennko]

Thanks to @cerebras, we now have access to high-quality OPEN SOURCE GPT models.

We've added them to xTuring - now you can fine-tune them using Alpaca dataset using LoRA in both 16b and 8b.

[Marktechpost AI @Marktechpost]

1/ Meet xTuring: An Open-Source Tool That Allows You to Create Your Own Large Language Model (LLMs) With Only Three Lines of Code
Quick Read: https://t.co/riW0RHjSYX
#ArtificialIntelligence @stochasticai

[Stochastic @stochasticai]

Using the Alpaca dataset, Stanford trained LLaMa to have ChatGPT-like performance, all under $500.

Recently, Databricks was able to do a similar job with Dolly by applying memory techniques, and on only 1 machine.

Now you can do the same w/ our xTuring and @cerebras