Rafael Oliveira

RAG vs. Long-Context LLMs I have yet to see a convincing paper or technical blog showing that long-context LLMs can or will replace RAG. So far I've seen specific long-context applications where long-context LLMs thrive and current retrieval benchmarks are not convincing. This new paper reports that longer-context LLMs suffer from diminished focus on relevant information, which is one of the primary issues that a RAG system addresses (i.e., uses more relevant information). They propose an order-preserving RAG mechanism that improves performance on long-context question answering. It's not perfect and in fact as retrieved chunks increases the quality of responses go up and then declines. But there is a sweet spot where it can achieve better quality with a lot fewer tokens than long-context LLMs. I am getting better results for RAG plus the added benefit of efficiency as well. Things could change rapidly but still not counting out RAG for now.

had the opposite of the intended effect in the end, teams run by non coding scrum masters and SAFe scaled agile consultants mandating all teams have the same practices.

After I published my guide to making slides I got a lot of great feedback & especially loved that others had already used it for their talks 🤩 In part 2, I'm showcasing examples from @t_redactyl and @vboykis and dive deeper into aesthetics & process. ines.io/blog/beautiful…

New open source OCR model just dropped! This one by Microsoft features the best text recognition I've seen in any open model and performs admirably on handwriting. It also handles a diverse range of vision tasks. You can play with it here: huggingface.co/spaces/gokaygo…

Branding. In the long run, the only unfair advantage that companies have over their competitors is their brands. In the long run, any advantage in technology disappears, and every company is competing to hire from the same pool of engineers to build things. If you have a good brand, you can earn high profit margins, even if what you produce is the same as your competitors. People would rather pay for something they can trust rather than go with something free and unreliable. That's why Apple and nVidia are the most valuable companies in the world. Take nVidia for example: other companies make hardware that's just as good, including Intel and AMD. If you are going to invest $10 million in building a data center to train models, you don't nVidia. But because of branding, you choose to pay a premium for nVidia anyway, making the company the most valuable in the world. AI companies hope to achieve the same thing: make such a solid brand that people will pay them a premium eventually.

@irace @ananayarora He's probably afraid his poor silly @grok is dead-on-arrival once LLMs get OS-level integration

@ananayarora Did you expect anything different

I can't believe this guy calls himself an "engineer" without understanding that it's on device and opt-out. The most hypocritical part of this is that Tesla captures so much more data everywhere you drive

@GossiTheDog I don't get how Microsoft released such a bad implementation. They even have tools to detect PII like github.com/microsoft/pres… that would have improved some of this

@goenning @Hetzner_Online is awesome, that's how

16 vCPU + 32 GB RAM + 20 TB Bandwith for €29!!! WHAT THE ACTUAL F? HOW?

@dvassallo @tdinh_me Reminds me when we hired someone in the UK. To pay taxes you need to have a bank account, and the banks we spoke to would not let us open an account without 1M£ in revenue in the UK, and flying there for an in person meeting 🤦‍♂️

@tdinh_me Same me 5yrs ago:

Brex says I'm too broke to have an account with them 🥲

@Pauline_Cx Now LinkedIn: 100% the first one 😅

X DMs: - 45%: "Hey, do you want to increase your MRR by +60% in 3 weeks?" - 45%: "I love you product. Btw I'm launching on PH, can you upvote?" - 25%: "Hi" - 5%: Genuine question, kind message

To qualify as Science a piece of research must be correct and reproducible. To be correct and reproducible, it must be described in sufficient details in a publication. To be 'published' (to receive a seal of approval) the publication must be checked for correctness by reviewers. To be reproduced, the publication must be widely available to the community and sufficiently interesting. If you do research and don't publish, it's not Science. Without peer review and reproducibility, chances are your methodology was flawed and you fooled yourself into thinking you did something great. No one will ever hear about your work. No one will pick it up and build on top of it. No one will build new technology and products with it. Your work will have been in vain. You'll die bitter and forgotten. If you never published your research but somehow developed it into a product, you might die rich. But you'll still be a bit bitter and largely forgotten.

@davidfowl @moojii Also for some reason totals broken on macOS, tried loading a simple project and couldn't figure out why it wouldn't, something like not finding the dotnet sdk despite it being installed and on the path

@moojii Can you record that experience and file an issue here github.com/dotnet/vscode-…

As someone who has built up muscle memory for Visual Studio over the last 15+ years, I've been trying to get accustomed to VS code. Some of the paradigms are beginning to click. I've been using it daily to work on some bigger projects.

@theojaffee @nearcyan Yeah, maybe? I first heard about it from @deepfates quoting this tweet from @allgarbled twitter.com/allgarbled/sta…

Slop is the new name for unwanted AI-generated content simonwillison.net/2024/May/8/slo…

@GroqInc @SHACK15sf @cerebral_valley @AIatMeta @csTimSears No remote teams?

@SHACK15sf @cerebral_valley @AIatMeta @csTimSears To register: partiful.com/e/p5bNF0WkDd1n…

We'll be at the Llama 3 hackathon at @SHACK15sf along with @cerebral_valley,@AIatMeta and all of the other partners. Stop by our table to chat, come hear @csTimSears's speaker workshop, and DM us for increased rate limits during the hackathon. What can you build on Groq?

It no longer lies

Microsoft CTO: "Thoughts on OpenAI" June 12, 2019

Introducing ChatQA-1.5, a family of models that surpasses GPT-4-0613 and Command-R-Plus on RAG and conversational QA. ChatQA-1.5 has two variants: Llama3-ChatQA-1.5-8B, huggingface.co/nvidia/Llama3-… Llama3-ChatQA-1.5-70B, huggingface.co/nvidia/Llama3-… We also open source our instruction tuning data, offer the ChatRAG Bench for evaluation, and provide a fine-tuned multi-turn QA retriever for the community. For more details, visit the following links: huggingface.co/collections/nv…

In the last 5 years I've seen a good number of solo founders go from $0 to $10k/month. Some of them are friends, and I know how much they struggled in the beginning. None of them had huge launches. They had been stuck below $1k/month for a while. Their initial metrics were not great. But contrary to the popular narrative to fail fast and keep churning out new products until you hit big, they stuck to it. One small marketing win at a time for weeks, months, and years until they got there.