sert_121

I wanted one place to read people who offer interesting perspectives and research about algorithms, compute and the emerging capabilities of LLMs.

So I manually picked about 210 candidate blogs from HN, Lobsters, GitHub Pages, Bear blogs etc. I went through them yes/no based on personal preferences and kept 77. Some are well known (Lil'Log, Interconnects, SemiAnalysis, Epoch). Many are small personal blogs that post a few times a year but are worth reading when they do.

Here's the website: findingfishes.vercel.app & One merged RSS feed (if you still use them (: ): findingfishes.vercel.app/feed.xml

I'd love suggestions for any blogs I missed!

I feel a lot of people have already addressed the broad stretches to which people are interpreting this. Regardless, the blogpost does open interesting directions that could be useful down the line (2-5years): a) tapping into the embedding space in ways that we don't, ie editing model weights to alter task-performance (mostly preventing degradation) b) Exploring the weight space in more intuitive ways c) Understanding how we can ablate on different parts of the model to understand neuron-based interp better

Seeing this as a replacement of training seems naive, yet one can see this feedback loop targeting weights as an viable alternative to let's say doing LoRA on certain models, especially if we don't want to teach new behaviours.

personal website: sert121.vercel.app rambles: https://sert121.bearblog.dev/

This jogged some memories! Was working on something that used radio waves to detect objects and humans ~5 years ago, I see that we've come a long way since then.

One of our goals(abandoned) was to also extend to wifi routers, so I am excited to see continued interest in this space!

https://www.sensorsportal.com/HTML/ST_JOURNAL/PDF_Files/P_32...

I am sure it already knows a lot regardless of the memory feature, as long you're sharing your chat history/ have your history enabled, but I agree, it'd simply worsen it.

I see your point in user trust, and that's fair, but the same concerns have been prevalent since GPT3 rolled out , that no one would trust these tools to write or edit anything. However, since then users are growing to be more and more attuned to filter and distinguish the quality of the responses (doing invisible A/B testing of these responses), so maybe that's what providers want to capitalize on.

ChatGPT has become one of the top-most browsed websites, and they want to capitalize on it even if 2% of the people actually trust the new integrations.

I agree with the fact that a monopolized web is not friendlier to anyone. But seeing the trajectories of tech companies in the past decade, the unfortunate north star is distribution and the relentless pursuit of it.

Even in that dire circumstance, I wish that the web versions keep up/are maintained, instead of being slowly deprecated, which happened for a lot of mobile-native versions of applications.

It's interesting to see how chatgpt is becoming more and more of a starting point of the web exploration, at which they're like, why even bother searching at this point, we'll just have default workflows for maps, buy (integration of stripe already marks it), booking airlines etc, which covers so much basic stuff people would do anyways.

The biggest bottleneck for this for the past two years imo wasn't the models, but the engineering and infra around it, and the willingness of companies to work with openaio directly. Now that they've grown and have a decent userbase, companies are much more willing to pay/or involve themselves in these efforts.

This has eventual implications outside user-heavy internet use (once we see more things built on the SDK), where we're gonna see a fork in the web traffic of human centric workflows through chat, and an seo-filled, chat/agent-optimized web that is only catered to agents. (crossposted)

I believe its a mix of three factors, (a) lack of transfer of institutional knowledge (b) lesser fundamental incentives for people to get better at fundamental skills/gaps (c) rise in hotfixes as we deal with time/scales that operate much faster, burn faster, and want to expand faster.

All of the above is multiplied 1.3x-1.5x with accelerating ways to get upto speed with iterative indexing of knowledge with llms. I believe we are reliant on those early engineers whose software took a while to build (like a marathon), and not short-sprinted recyclable software we keep shipping on it. The difference is not a lot of people want to be in those shoes (responsibility/comp tradeoffs.

is this supposed to be feed only for ambigrams? or more of a contest?

Hi, I've been trying to expose the llama.cpp as an API, you mentioned that found llamacpphtmld to solve this problem, can you share the link to this project? couldn't find it over github. Thanks!