Previously, this existed as crixet.com [0]. At some point it used WASM for client-side compilation, and later transitioned to server-side rendering [1][2]. It now appears that there will be no option to disable AI [3]. I hope the core features remain available and won’t be artificially restricted. Compared to Overleaf, there were fewer service limitations: it was possible to compile more complex documents, share projects more freely, and even do so without registration.
On the other hand, Overleaf appears to be open source and at least partially self-hostable, so it’s possible some of these ideas or features will be adopted there over time. Alternatively, someone might eventually manage to move a more complete LaTeX toolchain into WASM.
I was using Crixet before I switched over to Typst[0] for all of my writing. However, back when I did use Crixet, I never used its AI features. It was just a much better alternative to Overleaf for me. Sad to see that AI will be forced on all Crixet users now.
Very unfortunately named. OpenAI probably (and likely correctly) estimated that 13 years is enough time after the Snowden leaks to use "prism" for a product but, for me, the word is permanently tainted.
This seems like a very basic overleaf alternative with few of its features, plus a shallow ChatGPT wrapper. Certainly can’t compete with using VS Code or TeXstudio locally, collaborating through GitHub, and getting AI assistance from Claude Code or Codex.
> Certainly can’t compete with using VS Code or TeXstudio locally, collaborating through GitHub, and getting AI assistance from Claude Code or Codex.
I have a phd in economics. Most researchers in that field have never even heard of any of those tools. Maybe LaTeX, but few actually use it. I was one of very few people in my department using Zotero to manage my bibliography, most did that manually.
The quality and usefulness of it aside, the primary question is: are they still collecting chats for training data? If so, it limits how comfortable, and sometimes even permitted, people would with working on their yet-to-be-public work using this tool.
The collect chat records for any number of users, not the least of which being NSA surveillance and analysis - highly likely given what we know from the Snowden leaks.
With a tool like this, you could imagine an end-to-end service for restoring and modernizing old scientific books and papers: digitization, cleanup, LaTeX reformatting, collaborative or volunteer-driven workflows, OCR (like Mathpix), and side-by-side comparison with the original. That would be useful.
From my perspective as a journal editor and a reviewer these kinds of tools cause many more problems than they actually solve. They make the 'barrier to entry' for submitting vibed semi-plausible journal articles much lower, which I understand some may see as a benefit. The drawback is that scientific editors and reviewers provide those services for free, as a community benefit. One example was a submission their undergraduate affiliation (in accounting) to submit a paper on cosmology, entirely vibe-coded and vibe-written. This just wastes our (already stretched) time. A significant fraction of submissions are now vibe-written and come from folks who are looking to 'boost' their CV (even having a 'submitted' publication is seen as a benefit), which is really not the point of these journals at all.
I'm not sure I'm convinced of the benefit of lowering the barrier to entry to scientific publishing. The hard part always has been, and always will be, understanding the research context (what's been published before) and producing novel and interesting work (the underlying research). Connecting this together in a paper is indeed a challenge, and a skill that must be developed, but is really a minimal part of the process.
> these kinds of tools cause many more problems than they actually solve
For whom? For OpenAI these tools are definitely the solutions. They are developing by throwing various AI-powered stuff at the wall to see what sticks. These tools also demonstrate to the investors that innovation did not stall and to show that AI usage is growing.
Same with Microsoft: none of the AI stuff they are shoving down the users' throats were actually designed for the users. All this stuff is only for the token usage to grow for the shareholders to see.
Similar with Google although no one can deny real innovation happening there.
The real problem is that researchers are pushed to publish as their publication is the only way their career can advance. It's not even to "boost" your CV, as a researcher your publication history IS your CV.
It was already a problem 25 years ago when I did my Ph.D., and I don't think things changed that much since then.
This encourages researchers to publish barely valuable results, or to cut one articles into multiple ones with small variations to increase their number of publications. Also publishers creating more conferences and more journals to respond to the need that researchers have to publish.
I remember many experienced professors telling me cynically about this, about all the techniques they had to blow up one small finding into many articles.
Anyway - research slop started way before AI. It's probably going to make the problem worse, but the root issue have been there for a long time.
in the end we're going to end up with papers written by AI, proofread by AI .....summarized for readers by AI. I think this is just for them to remain relevant and be seen as still pushing something out
It's interesting how quickly the quest for the "Everything AI" has shifted. It's much more efficient to build use-case specific LLMs that can solve a limited set of problems much more deeply than one that tries to do everything well.
I've noticed this already with Claude. Claude is so good at code and technical questions... but frankly it's unimpressive at nearly anything else I have asked it to do. Anthropic would probably be better off putting all of their eggs in that one basket that they are good at.
All the more reason that the quest for AGI is a pipe dream. The future is going to be very divergent AI/LLM applications - each marketed and developed around a specific target audience, and priced respectively according to value.
In my lab, we have been struggling with automated image segmentation for years. 3 years ago, I started learning ML and the task is pretty standard, so there are a lot of solution.
In 3 months, I managed to get a working solution, which only took a lot of sweat annotating images first.
I think this is where tools like OpenCode really shine, because they unlock the potential for any user to generate a solution to their specific problem.
Not an academic, but I used LaTeX for years and it doesn’t feel like what future of publishing should use. It’s finicky and takes so much markup to do simple things. A lab manager once told me about a study that people who used MS Word to typeset were more productive, and I can see that…
This does way less than i'd expect. Converting images to tikz is nice but some of the other applications demonstrated were horrible. This is no way anyone should be using AI to cite.
Anybody else notice that half the video was just finding papers to decorate the bibliography with? Not like "find me more papers I should read and consider", but "find papers that are relevant that I should cite--okay, just add those".
I’ve been “testing” LLM willingness to explore novel ideas/hypotheses for a few random topics[0].
The earlier LLMs were interesting, in that their sycophantic nature eagerly agreed, often lacking criticality.
After reducing said sycophancy, I’ve found that certain LLMs are much more unwilling (especially the reasoning models) to move past the “known” science[1].
I’m curious to see how/if we can strike the right balance with an LLM focused on scientific exploration.
[0]Sediment lubrication due to organic material in specific subduction zones, potential algorithmic basis for colony collapse disorder, potential to evolve anthropomorphic kiwis, etc.
[1]Caveat, it’s very easy for me to tell when an LLM is “off-the-rails” on a topic I know a lot about, much less so, and much more dangerous, for these “tests” where I’m certainly no expert.
Way too much work having AI generate slop which gets dumped on a human reviewer to deal with. Maybe switch some of that effort into making better review tools.
129 comments
[ 3.7 ms ] story [ 101 ms ] threadOn the other hand, Overleaf appears to be open source and at least partially self-hostable, so it’s possible some of these ideas or features will be adopted there over time. Alternatively, someone might eventually manage to move a more complete LaTeX toolchain into WASM.
[0] https://crixet.com
[1] https://www.reddit.com/r/Crixet/comments/1ptj9k9/comment/nvh...
[2] https://news.ycombinator.com/item?id=42009254
[3] https://news.ycombinator.com/item?id=46394937
[0]: https://typst.app
It also offers LaTeX workspaces
see video: https://www.youtube.com/watch?v=feWZByHoViw
I have a phd in economics. Most researchers in that field have never even heard of any of those tools. Maybe LaTeX, but few actually use it. I was one of very few people in my department using Zotero to manage my bibliography, most did that manually.
I thought this was introduced by the NSA some time ago.
Uhm ... no.
I think we need to put an end to AI as it is currently used (not all of it but most of it).
Was this not already possible in the web ui or through a vscode-like editor?
The collect chat records for any number of users, not the least of which being NSA surveillance and analysis - highly likely given what we know from the Snowden leaks.
There was an idea of OpenAI charging commission or royalties on new discoveries.
What kind of researcher wants to potentially lose, or get caught up in legal issues because of a free ChatGPT wrapper, or am I missing something?
I'm not sure I'm convinced of the benefit of lowering the barrier to entry to scientific publishing. The hard part always has been, and always will be, understanding the research context (what's been published before) and producing novel and interesting work (the underlying research). Connecting this together in a paper is indeed a challenge, and a skill that must be developed, but is really a minimal part of the process.
For whom? For OpenAI these tools are definitely the solutions. They are developing by throwing various AI-powered stuff at the wall to see what sticks. These tools also demonstrate to the investors that innovation did not stall and to show that AI usage is growing.
Same with Microsoft: none of the AI stuff they are shoving down the users' throats were actually designed for the users. All this stuff is only for the token usage to grow for the shareholders to see.
Similar with Google although no one can deny real innovation happening there.
It was already a problem 25 years ago when I did my Ph.D., and I don't think things changed that much since then.
This encourages researchers to publish barely valuable results, or to cut one articles into multiple ones with small variations to increase their number of publications. Also publishers creating more conferences and more journals to respond to the need that researchers have to publish.
I remember many experienced professors telling me cynically about this, about all the techniques they had to blow up one small finding into many articles.
Anyway - research slop started way before AI. It's probably going to make the problem worse, but the root issue have been there for a long time.
I've noticed this already with Claude. Claude is so good at code and technical questions... but frankly it's unimpressive at nearly anything else I have asked it to do. Anthropic would probably be better off putting all of their eggs in that one basket that they are good at.
All the more reason that the quest for AGI is a pipe dream. The future is going to be very divergent AI/LLM applications - each marketed and developed around a specific target audience, and priced respectively according to value.
In my lab, we have been struggling with automated image segmentation for years. 3 years ago, I started learning ML and the task is pretty standard, so there are a lot of solution.
In 3 months, I managed to get a working solution, which only took a lot of sweat annotating images first.
I think this is where tools like OpenCode really shine, because they unlock the potential for any user to generate a solution to their specific problem.
This is all pageantry.
The earlier LLMs were interesting, in that their sycophantic nature eagerly agreed, often lacking criticality.
After reducing said sycophancy, I’ve found that certain LLMs are much more unwilling (especially the reasoning models) to move past the “known” science[1].
I’m curious to see how/if we can strike the right balance with an LLM focused on scientific exploration.
[0]Sediment lubrication due to organic material in specific subduction zones, potential algorithmic basis for colony collapse disorder, potential to evolve anthropomorphic kiwis, etc.
[1]Caveat, it’s very easy for me to tell when an LLM is “off-the-rails” on a topic I know a lot about, much less so, and much more dangerous, for these “tests” where I’m certainly no expert.
Seems like they have only announced products since and no new model trained from scratch. Are they still having pre-training issues?
I can't wait