For experimentation and token-minimizing I have been running some open-weight LLMs (e.g., llama3.2 for text and qwen3-coder:30b for code) locally on a MacBook via ollama within VS Code. I noticed that long-context tasks (e.g., analyzing a complex codebase and generating a new feature) can take a long time and compute, eventually heating up the device to a hazardous degree, especially under an ambient room temperature over 90°F.
https://www.instagram.com/p/DaXY3tDj8nM/
I created this little code repo as a fun exercise of LangGraph for a news agency that generates fictional news articles based on real events. Some of these articles are available on this Substack site.
After having a meeting with the patent attorneys, usually you will receive two documents to review – a Word document with texts and a PDF document with drawings.
Aside from using LLM to summarize the former, a good strategy to review the documents (without falling asleep) is to look at the drawings for a visual overview, and then read the high-level sections, like the abstract, the background, and the claims, before the detailed description when necessary.
You might find the legalese hard to understand even though they are supposed to describe your invention. This is expected, as the the patent is often worded as a defense against infringement, so just make sure that the claims are broad enough to cover your invention but not too broad as to be non-patentable.
With SIGGRAPH, TVCG, NSF, and CHI I ended up reviewing 30+ submissions this month. It was quite some extra work, especially for an industry researcher who doesn’t have to serve on academic review committees. But I am glad that I would get (forced) to read topics that I usually would not, and really appreciate those who accepted my review invitations.
I am also surprised by the reluctance of some to review papers, especially for those who are in the academia whose careers depend on getting their own papers accepted and grants funded. If you are too busy to review other people’s submissions, you probably shouldn’t submit to begin with. Who should review yours then?
Most kids don’t like math, but this one managed to love it enough to self-teach advanced topics and eventually disprove a long-standing conjecture in harmonic analysis while still in high-school.
She treats math as art and it reflects on the way she designs her presentations and probably also approaches problems.
Sharing a research work that is not yet published (not even a preprint on arXiv) can have several benefits, such as receiving early feedback, increasing visibility (especially for junior researchers), and attracting potential collaborators.
On the other hand it also has some risks, such as being scooped by others or jeopardizing anonymity during the review process or public disclosure for patent application.
To maximize the benefits and minimize the risks, I suggest sharing only with trusted colleagues who have the expertise to provide constructive feedback and ethics to maintain confidentiality. If the paper is still under review, check the conference/journal policies (e.g., this for SIGGRAPH). For patent application, any public disclosure will invalidate international filings but there is a one-year window for US filings.
While at SIGGRAPH conferences I usually spent the majority of my “attendance time” in the animation festivals and production talks primarily because (1) the contents are usually not recorded for later consumption, (2) the effects look nice, and (3) the topics tend to be something I am less familiar with, even after 20+ years in the field.
The animation creations and effects productions are hybrids of art and science/engineering.
Each effect is tailored for the specific style and story and created by a combination of artistic practices and off-the-shelf/bespoke tools.
The presentations tend to focus on the design process and tools deployment, instead of technical methods and implementations.
Upon noticing a reminder email about upcoming reviews due for a conference 3 days ago, I realized that I haven’t received any prior notifications about the assignments and thus haven’t started reading the papers yet. It is not a good form to miss reviews, especially for a conference that I previously chaired, so I spent an intense long weekend reviewing these submissions.
Joining a hot research trend is like buying a stock when the price is high.
These neural network pioneers (like Hinton, LeCun, etc.) started their works not when the field was hot, but when it was considered a dead end. And they stuck with it for decades until the right circumstances (e.g., GPU compute finally became feasible). Now, every time a new idea (about generative X or whatever) pops in the head, there might already be 10 arXiv papers on it.
It is important to know any significant new advancements, but it is even more important to figure out what and how to apply these into our own research and be able to formulate our own (unique) vision.
Theme: Rubric. Get a free blog at WordPress.com