alt Hacker NewsMaking Wolfram tech available as a foundation tool for LLM systems
Comments
I agree, but to be truly foundational, it needs to be open source and accessible for everyone!
That’s why I’m working on an open source implementation of Mathematica (i.e. an Wolfram Language interpreter):
I know this marks me out as a very specific and fairly derided kind of nerd, but I’m really excited to see what new features will be included in Mathematica 15 which is presumed to be launching fairly soon.
the real value proposition here is correctness guarantees that LLMs fundamentally cant provide. when an LLM says 2+2=4 it arrived there statistically, not computationally. for anything safety-critical - engineering tolerances, drug dosage calculations, financial modeling - you want a deterministic engine producing the answer and the LLM just translating between human intent and formal queries.
the CAG framing is clever marketing but the underlying idea is sound: treat the LLM as a natural language interface to a computational kernel rather than the computation itself. weve been doing something similar with python subprocess calls from agent pipelines and it works well. the question is whether wolfram language offers enough over python+scipy+sympy to justify the licensing cost and ecosystem lock-in.
There's a great discussion with Stephen Wolfram on the Sean Carroll podcast. Listening to it made me think very highly of Wolfram. He's a free thinking, eccentric, mathematician, scientist; who got started doing serious work at a very young age. He still has a youthful creative approach to thought and science. I hope LLMs do pair well with his tools.
LLMs using code to answer questions is nothing new, it's why the "how many Rs in strawberry" question doesn't trip them up anymore, because they can write a few lines of Python to answer it, run that, and return the answer.
Mathematica / Wolfram Language as the basis for this isn't bad (it's arguably late), because it's a highly integrated system with, in theory, a lot of consistency. It should work well.
That said, has it been designed for sandboxing? A core requirement of this "CAG" is sandboxing requirements. Python isn't great for that, but it's possible due to the significant effort put in by many over years. Does Wolfram Language have that same level? As it's proprietary, it's at a disadvantage, as any sandboxing technology would have to be developed by Wolfram Research, not the community.
Either I miss something, or is it yet another marketing approach by Stephen Wolfram? The post talks about "foundation tool", yet it offers an MCP to a proprietary app (out of miriads of such).
Sure, as any other tech, Mathematica may have its edges (I used it deeply 10-15 years ago, before I migrated to Python/Jupyter Notebook ecosystem). But in the grand scheme of things, it is yet another tech, and one that is losing rather than gaining traction.
Certainly not "a new kind of science".
I like Mathematica and use it regularly. But I did not see any benefits of using it over python as a tool that Claude Code can use. Every script it produced in wolfram was slower with worse answers than python. Wolfram people are really trying but so far the results are not very good.
I tried using wolfram alpha as a tool for an llm research agent, and I couldn't find any tasks it could solve with it, that it couldn't solve with just Google and Python.
Recent shower thought: "Wolfram has a huge array of unique visualization tools for math. Wouldn't it be neat if LLM's could embed them in responses, for people wanting to teach themselves math that are visual learners?"
Every major technological invention nowadays quickly breeds open source clones that evolve to be on par with the commercial ones on some time scale. Why hasn't this happened to Wolfram Alpha/Mathematica? I know there's Sympy, but it's so far behind Mathematica that it's not even comparable. Is the heavily mathematical nature of the tool somehow an insurmountable obstacle to the open source community?
The blog post would have been more effective with a specific example of what it solves, a demo, or at least some anecdotes of what this has already solved via these integrations. As it stands, it comes off rather self-aggrandizing and a bit desperate, as though Wolfram tech perceives itself as threatened to remain relevant.
A simple skill markdown for Claude Code was enough to use the local Wolfram Kernel.
Even the documentation search is available:
```bash
/Applications/Wolfram.app/Contents/MacOS/WolframKernel -noprompt -run '
Needs["DocumentationSearch`"];
result = SearchDocumentation["query term"];
Print[Column[Take[result, UpTo[10]]]];
Exit[]'
```
Every time I go to purchase a hobby license for Wolfram / Mathematica I give up because their product offering is the most convoluted i've seen in my life.
Why can't I just pay some price and get the entire bundle of Wolfram One Cloud + API calls + LLM Assistant + This new MCP access + Mathematica?
I need to buy 5 different things - and how does that look for me the user, I need 5 different binaries?
They really should sort that out, I know they are losing money because of this. I emailed their support once and ended up getting more confused.
Having been both in academia and industry, to me Stephen always sounds too academic for business, but for academia, a mix of too "industry" and a crank.
Relevant discussion of Kagi and Wolfram (2024): https://news.ycombinator.com/item?id=39606048
Lots of big words there, but can I now expose the local Mathematica (confusingly renamed Wolfram a while ago) that I'm paying for, through MCP to Claude Code?
Because it seems I can't and all the big words are about buying something new.
Please check this book for machine learning and AI with reproducible codes in Wolfram language by Etienne Bernard [1].
[1] Introduction to Machine Learning:
https://www.wolfram.com/language/introduction-machine-learni...
One thing me.Stephen never made available is for people to copy results of Wolfram Alpha… he persisted doing so even after ocr and LLms were omnipresent, so somehow I don’t trust him even though his reload theory seems very appealing and apparently the team there understands production grammars very well since 1996.
Lines up with the current YC advice to "make something agents want". Not sure it makes a lot of sense to try and build a VC-backed business like this, but if distribution is the moat these days, perhaps.
Sounds cool.
Aside, I hate the fact that I read posts like these and just subconsciously start counting the em-dashes and the "it's not just [thing], it's [other thing]" phrasing. It makes me think it's just more AI.
Something this "shape" has been coalescencing since the first tool calls were done. To draw another Star Trek parallel, this reformulation is what Brent Spiner is during the little stares and pauses made before answering a complicated but constrained problems on the show. Onward!
I read his book “A new kind of science” and quickly figured out why it was self-published. My goodness it’s bad and need of an editor.
A big disappointment as I’m a fan of his technical work.
Maybe I’m not understanding but what is different than just using existing wolfram tools via an API? What is infinite about CAG?
I can't help but think about Wolfram every time I go into my thinking-about-ai mode. Really not sure how to frame all this stuff, but jeez there's a nexus, right?
Ultimately by making their language and tooling closed and paywalled they doomed it to never be relevant to an LLM ever.
Fucking finally. What took them so long. This needed to be done day one.
He's 10 years too late for that. That's how you lose by keeping your stuff proprietary. The world innovates without paying any attention to you and you get left behind.
Imagine if 10 years ago Wolfram software was opensourced. LLMs would be talking it since the day one.
>"But an approach that’s immediately and broadly applicable today—and for which we’re releasing several new products—is based on what we call
computation-augmented generation, or CAG.
The key idea of CAG is to inject in real time capabilities from our foundation tool into the stream of content that LLMs generate. In traditional retrieval-augmented generation, or RAG, one is injecting content that has been retrieved from existing documents.
CAG is like an infinite extension of RAG
, in which an infinite amount of content can be generated on the fly—using computation—to feed to an LLM."
We welcome CAG -- to the list of LLM-related technologies!
Imagine Isaac Newton (and/or Gottfried Leibniz) saying, "Today we're announcing the availability of new mathematical tools -- contact our marketing specialists now!"
The linked article isn't about mathematics, technology or human knowledge. It's about marketing. It can only exist in a kind of late-stage capitalism where enshittification is either present or imminent.
And I have to say ... Stephen Wolfram's compulsion to name things after himself, then offer them for sale, reminds me of ... someone else. Someone even more shamelessly self-promoting.
Newton didn't call his baby "Newton-tech", he called it Fluxions. Leibniz called his creation Calculus. It didn't occur to either of them to name their work after themselves. That would have been embarrassing and unseemly. But ... those were different times.
Imagine Jonas Salk naming his creation Salk-tech, then offering it for sale, at a time when 50,000 people were stricken with Polio every year. What a missed opportunity! What a sucker! (Salk gave his vaccine away, refusing the very idea of a patent.)
Right now it's hard to tell, but there's more to life than grabbing a brass ring.
CAG sounds like fake solution for LLM's. Math problems are not custom data, they are limited in amount, and do not refresh like product manuals.
Hence math can always be part either generic llm or math fine tuned llm, without weird layer made for human ( entire wolfram) and dependencies.
Wolfram alpha was always an extra translation layer between machine and human. LLM's are a universal translation layer that can also solve problems, verify etc.
My first thought on CAG was that it sounds a bit like bolting on an MCP server onto SymPy (AFAIK, the closest OSS thing to Mathematica). And it turns out someone has already done that.
There's a lot of value in the implementation of many strong and fast algeorithms in computer algebra in proprietary tools such as Maple, Wolfram, Matlab. However, I (though of course believe that such work needs to be compensated) find it against the spirit of science to keep them from the general public. I think it would be good service to use AI tools to bring open source alternatives like sympy and sage and macaulay to par. There's really A LOT of cool algorithms missing (most familiar to me are some in computational algebraic geometry)
Additionally I think because of how esoteric some algorithms are, they are not always implemented in the most efficient way for today's computers. It would be really nice to have better software written by strong software engineers who also understands the maths for mathematicians. I hope to see an application of AI here to bring more SoTA tools to mathematicians--I think it is much more value than formalization brings to be completely honest.