Blog

Solving the Yield Problem #

How we solved the yield problem

via Hacker News

This was a fun read, especially after already having a bit of background knowledge on chip manufacturing. Cerebras and Groq both utilize custom chips to power their ridiculously fast AI inference, but I hadn’t previously considered the widely differing order of magnitudes these chips have compared to traditional GPUs. As an example, Cerebras cites their WSE-3 as having 900,000 cores compared to individual H100s which have only 144.

Cerebras’ method for reducing the impact of defects feels quite straightforward—reducing core sizes and having a surplus of cores means that individual core defects are able to be had at an extremely small scale while having almost no impact on the overall chip. Additionally, Cerebras uses a special routing architecture to ensure that cores are able to route around disabled cores, meaning that disabling cores creates no interference. Through these methods, Cerebras notes the WSE-3 as having a high 93% core utilization, compared to H100s with a 92% utilization.

Ndea #

Ndea

via @fchollet.bsky.social

A new AI startup cofounded by François Chollet, the creator of Keras. Ndea’s main focus seems to be developing AGI, which is obviously a shared interest of all the major AI companies.

As described on their website, Ndea is utilizing guided program synthesis to reach AGI:

Instead of interpolating between data points in a continuous embedding space, program synthesis searches for discrete programs, or models, that perfectly explain observed data.

While their description of program synthesis is a bit overcomplicated, it seems to simply be searching for certain programs that interpret and output data in a correct way. Since this program search can be guided through deep learning, Ndea claims that programs can be found which perfectly model observed data without necessitating compromises.

It will be interesting to see if this startup will be able to make a large, independent impact on AI research, or if it will just be another endeavour that results in being acquired by a larger company.

o1 Thinking In Chinese #

o1 ‘thinks’ in Chinese

via Slashdot

While I haven’t personally encountered it in a while, ChatGPT chats used to rarely autogenerate titles in Spanish instead of English. I assume that o1 thinking in Chinese occurs for a similar set of reasons, since both seem related to the multilingual datasets that the models are trained on. As an aside, I’m curious if Chinese makes up a significantly higher percentage of the training data for Chinese LLMs like DeepSeek, since both English and Chinese already make up the significant part of the corpuses for English models.

DoomPDF #

doom.pdf

via Hacker News

The highly anticipated sequel to PDFTris. It’s surprisingly responsive for how limiting PDFs are, but it makes sense given that most of the logic is just JavaScript.

PDFtris #

pdftris.pdf

via Hacker News

Cool demo of the concerning program execution that is built into PDFs. I initially assumed that PostScript was used to add logic, but it apparently instead uses JavaScript, which I didn’t know was able to be embedded into PDFs. Note that it doesn’t seem to work in Preview.app, but it does work in the builtin PDF viewers in Chrome and Firefox.