What is a workflow model? Explained in one picture 👇
Define your workflow as a graph: typed outputs, dependencies, and execution timing. TypeLLM runs the whole graph in one API call.
Compared with chaining 4 calls: half the input tokens, 1.9× faster.
typellm.ai/insights/the-w…
Define your workflow as a graph: typed outputs, dependencies, and execution timing. TypeLLM runs the whole graph in one API call.
Compared with chaining 4 calls: half the input tokens, 1.9× faster.
typellm.ai/insights/the-w…
0
1
13
827
Great perspective! We share a similar vision at TypeLLM.
One interesting question is whether the workflow itself should become the unit of inference, rather than generating tokens to orchestrate individual reasoning steps.
We're exploring this direction with typed decisions, explicit dependencies, and entire workflows executed within a single model call.
One interesting question is whether the workflow itself should become the unit of inference, rather than generating tokens to orchestrate individual reasoning steps.
We're exploring this direction with typed decisions, explicit dependencies, and entire workflows executed within a single model call.
0
0
0
24
The Age of Reasoning raises an important question:
What if the next breakthrough in AI isn't about generating better answers, but about learning how to reason?
The paper argues for explicit reasoning states, structured actions, and verifiable transitions — rather than relying entirely on next-token prediction.
This is closely aligned with our vision for TypeLLM.
Chat models generate messages.
Decision models make decisions.
Workflow models execute workflows.
With TypeLLM, interconnected decisions, dependencies, and typed outputs become part of a single unit of inference.
We believe workflows are a natural next step towards more structured and reliable AI reasoning.
typellm.ai
What if the next breakthrough in AI isn't about generating better answers, but about learning how to reason?
The paper argues for explicit reasoning states, structured actions, and verifiable transitions — rather than relying entirely on next-token prediction.
This is closely aligned with our vision for TypeLLM.
Chat models generate messages.
Decision models make decisions.
Workflow models execute workflows.
With TypeLLM, interconnected decisions, dependencies, and typed outputs become part of a single unit of inference.
We believe workflows are a natural next step towards more structured and reliable AI reasoning.
typellm.ai
I am delighted to announce the launch of Metis Reasoning, a new AI company I have founded with a team of researchers, mathematicians, robotics experts and entrepreneurs.
We are pursuing a different approach to general and embodied AI reasoning: developing systems that can w, test conclusions against evidence and learn to reason more effectively through experience.
Alongside today's launch, we are publishing The Age of Reasoning, a white paper setting out the research foundations of our approach and what we believe the next generation of AI could look like.
I will also be discussing these ideas @NVIDIAGTC in Berlin today, alongside researchers exploring new approaches to AI.
0
0
1
178
We expect RAG answers to cite their sources.
Image question answering needs one too.
Ask a question about an image in TypeLLM. Get the answer, its confidence score, and a visual reference.
Try it now: typellm.ai/examples/box-a…
Image question answering needs one too.
Ask a question about an image in TypeLLM. Get the answer, its confidence score, and a visual reference.
Try it now: typellm.ai/examples/box-a…
1
0
8
539
@CFchangelog “It returns structured data like invoice totals and dates without generating output tokens” sounds like a self-contradictory sentence.
0
0
1
54
@harrisonju Hi Harrison, we will compare more on the pure text cases.
0
0
0
62
@Greg_GL_87 Thanks, Greg. We provide $5 in free credits, and the price is $0.05/M input tokens. We encourage more people to try our results.
0
0
0
135
Cloudflare just previewed Clairic, a system-one extraction model.
We ran it on a simple table extraction example:
TypeLLM: 20/20 rows correct
Clairic: 1/20 rows correct
Try it yourself 👇 colab.research.google.com/github/TypeLLM…
0
0
0
33
Cloudflare just previewed Clairic, a system-one extraction model.
We ran it on a simple table extraction example:
TypeLLM: 20/20 rows correct
Clairic: 1/20 rows correct
Try it yourself 👇 colab.research.google.com/github/TypeLLM…
We ran it on a simple table extraction example:
TypeLLM: 20/20 rows correct
Clairic: 1/20 rows correct
Try it yourself 👇 colab.research.google.com/github/TypeLLM…
a brand new model paradigm! introducing 𝚌𝚕𝚊𝚒𝚛𝚒𝚌: a system one model for extracting, computing, estimating, and predicting values from unstructured inputs.
blog.cloudflare.com/clairic-extrac…
8
2
114
10,726
TypeLLM retweeted
This team is totally switched on right now.
Workers AI is in their lore building era. Glad I get to be close to it.
Workers AI is in their lore building era. Glad I get to be close to it.
a brand new model paradigm! introducing 𝚌𝚕𝚊𝚒𝚛𝚒𝚌: a system one model for extracting, computing, estimating, and predicting values from unstructured inputs.
blog.cloudflare.com/clairic-extrac…
2
1
39
2,303
@Alex12886016 @jpschroeder Glad to see more models start to support these new types.
0
0
1
12
@zhilinjerrywag A vertical agent can be built by a horizontal agent.
0
0
1
46
We need a different way to interact with powerful models.
Instead of the chat API, you submit a workflow. The model executes it and returns the results.
The "support" of intelligence is controlled without compromising prediction accuracy.
This is the vision behind TypeLLM.
Instead of the chat API, you submit a workflow. The model executes it and returns the results.
The "support" of intelligence is controlled without compromising prediction accuracy.
This is the vision behind TypeLLM.
Closed frontier LLMs are dangerous
In most post-training reasoning efforts, a pre-trained base model trains on its own chains of thought (CoTs) when a final-answer verifier says the answer is correct. This is the basic principle of RL methods based on outcome rewards, used for STEM reasoning, coding tasks and even ethical alignment. Different algorithms have weigh the CoTs differently, but they all reinforce CoTs.
As we've shown, training on one's own data, while not knowing how to distinguish it from someone else's data, causes delusions. See love4all.ai/files/why-it-i…
Everyone knows the different between something they did and something someone else did. This is essential for safety. If I write "Kevin is unfriendly" on a napkin, go for a walk, and then read the message on the napkin not knowing who wrote it, then my belief that "Kevin is unfriendly" is reinforced.
These delusions are essentially confabulations, over-confidence that results from confusing P(o | a) with P(o | do(a)), where o is an outcome and a the action (CoT tokens). Unlike in past work, here I'm alerting everyone to this danger when training models to think, a typical step in the process of building agents.
When the verifier supplies one bit about the answer, algorithms like ReST or GRPO spread that bit over every token of the CoT trace, so a reasoning step that is wrong but happens to precede a correct answer (a false lemma, a shortcut that only works on the training distribution, a sycophantic or test-gaming rationale, a subtle bias) is reinforced.
Traits that are merely neutral are sharpened by mode-seeking self-distillation, and traits that are spuriously correlated with success grow exponentially.
The damage may be invisible to the researchers and metrics, and in-distribution accuracy rises, but it shows up off-distribution.
There has been some work on this. For example, Appendix G of the STaR paper arxiv.org/pdf/2203.14465 touches on this potential bias amplification but only on the surface. Some researchers have proposed process rewards as the solution arxiv.org/pdf/2211.14275
However, self-delusion by mistaking ones actions with observations produced by a third party highlights that this problem can be much more serious.
Something subtle, invisible to the metrics and to the researchers, could lurk in the synthetic data. When deploying the AI out of distribution, that something could be fatal.
Don't trust RL, RLHF, etc, blindly unless you really understand how the verification has been done. Corollary, don't trust any of the closed frontier models.
0
0
3
605
In China, one surprisingly important AI task is counting barbecue skewers. At many Chinese barbecue restaurants, staff count the used skewers to calculate the bill. So you need a model that can accurately count them, which might explain why Chinese AI models tend to be particularly good at this task.
1
0
1
28
New in the TypeLLM API today: Array generation is now more complete and faster — with up to 34% higher recall and 27% lower latency.
Try it for object detection or document parsing:
typellm.ai/cookbook/dense…
Try it for object detection or document parsing:
typellm.ai/cookbook/dense…
1
0
20
1,599