Yousuf Jawwad retweeted
Google Jeff Dean says bigger context windows alone are not enough
What matters is staged retrieval: lightweight mechanisms that narrow a trillion tokens down to 10 million, then to the million you actually need
"you don't need a trillion at once, you need the right million"
Yousuf Jawwad retweeted
The most common mistake young founders make is forcing everyone to work 24/7 or 996.
Here's 5 reasons why:
– Very talented people, paradoxically, will work hard mimetically when they see people around them doing it, without the burden of compulsion. They want a carrot, not a stick.
– It alienates practically everyone >35 or with kids. Experienced people who can save you months of time in the future making good decisions in the present, not just blind velocity. The best engineers can get more done in 40hrs/wk than most in 80.
– The cost of losing someone good and familiar with your stack to burnout is far greater than the +20% hours. Two quarters of missed revenue, and it will hit your retention numbers.
– I get it. It takes hard work to will a startup into existence. Founders should put in the effort (you, after all, have 25-100x the equity of your employees) and set the bar for what's expected. Reward those who meet the bar, remove those who don't.
– Freeing up the compulsion to work all the time allows the best people to have fantastic ideas. Ideas need time to breathe and blossom.
There are always exceptions, but generally you don't build generational companies with a slave-driver mentality. Culture is important. You want your employees to love their job and tell their friends that. Early engineers at companies like Google, OpenAI and Anthropic worked because they wanted to, not because they were forced to.
Build a culture where people want to work hard. Don't force them to.
Yousuf Jawwad retweeted
The rivalry between Knuth and Douglas McIlroy is well known. While McIlroy favoured real-world implementation, Knuth was known for mathematical and algorithmic elegance.
When asked to implement a program to list the most frequent words in a file, Knuth wrote a 10-page program using sophisticated data structures like tries and following literate programming.
While McIlroy wrote a six line shell script that leveraged unix pipelines.
It was widely debated which approach was better.
This rivalry also shows up in a paper by McIlroy and Jon Bentley where they improved quick sort for real-world inputs. At one place you can see a comment on Knuth.
They note how Knuth downplayed the performance impact of swapping and comparison operations as if they are about the same as ordinary bookkeeping operations. But when measured for real-world data that is just not integers, these operations were 3x-9x more expensive than bookkeeping operations.
Yousuf Jawwad retweeted
Cassandra Team at Apple is searching for a fresh grad / person early in their career to join our ranks in SF/Bay Area!
Come work on super interesting problems with world class team. Help us build better Cassandra!
Ping me if you’re interested!
jobs.apple.com/en-us/details…
Yousuf Jawwad retweeted
Saas startup founder: “We’ve found that applicants with recently created LinkedIn profiles (last 6-12 months) with several YOE are almost always fake. As in 9/10 fake. So we are now immediately rejecting these.
We might lose on some genuine applicants this way, but fake profiles are out of hand.”
Fake applicants are really a big problem…
Yousuf Jawwad retweeted
in particular, if you have thought about how to squeeze max performance out of a system, we'd love to talk to you.
and if you have a background in compiler design or programming language design, we might have something great for you.
Yousuf Jawwad retweeted
Rippling sued @Deel today. Our lawsuit alleges Deel cultivated a spy at Rippling & orchestrated a long-running trade-secret theft. The spy searched “deel” in our systems 23 times per day on avg, letting him spy on Deel’s own customers who were considering a switch to Rippling.
Yousuf Jawwad retweeted
Just like we containerized webapps, we'll have to containerize AI agents... but how? What first principles led to containers in the first place, and how do they transpose to LLMs?
This was my first talk at an AI meetup, but I doubt it will be the last!
Recording 👇
Yousuf Jawwad retweeted
1/14 Go's new map implementation in 1.24 is powered by Swiss Tables, a cutting-edge hash table design that significantly boosts performance. Let's break down why this matters and how @CockroachDB's implementation played a key role. 🧐🚀
Yousuf Jawwad retweeted
~12yrs ago, I got a job @Google. Those were still early days of cloud. I joined GCP @<150M ARR & left @~4B (excld GSuite). Learned from some of the smartest ppl in tech. But we also got a LOT wrong that took yrs to fix. Much of it now public, but here’s my ring-side view👇
Yousuf Jawwad retweeted
We're open-sourcing an internal tool that may be useful to other NixOS users. But even if you don't want to embrace Madness, you may enjoy the story of why we wrote it, the lessons we learned, and the terrible cost to our sanity: antithesis.com/blog/madness/
Yousuf Jawwad retweeted
There’s a lot of stuff going on in AI right now and I wanted to share a few thoughts:
1️⃣ This is the worst that ChatGPT will ever be, and if @sama is saying their model improvements are actually accelerating, then everyone’s thoughtful timescales for when “AI as powerful as anyone can reasonably need” all gotta be halved. It’s all getting here faster.
2️⃣ Google just cracked infinite context windows earlier this month. Context windows only bounded by MEMORY. Once implemented and available I think it will megaaccelerate Fortune 500 adoption of AI as they can shovel every morsel of their business data in and get back something even more interesting than what Deloitte just built for them. A second corporate AI boom.
3️⃣ Groq (@GroqInc) has built a unique hardware-software layer that can run models in the cloud faster than any other company in the space, and it’s not close. On Reddit, a Groq engineer said, “We built a deterministic compute fabric (read: lots of chips working together synchronously) that uses this fact to plan the entire inference ahead of time.” Groq saw an opportunity in the market to focus on inference (running the model, running queries) in a way that nobody else can. This allows Groq to run Llama 3 (70B) at 284 tokens/second. For comparison, ChatGPT 4 in your browser is about 15 tokens/second on its best day. Devs building AI into their apps only have 2 smart options now: calling gpt-4 (OpenAI) or Llama 3 (Groq). Why would you use anything else?
4️⃣ Perplexity (@perplexity_ai) is the best tool for general purpose usage of AI in my daily life. If I need to learn about something, I ask Perplexity and not Google. It’s just so much better that there’s actually not a single downside. It’s faster and better in every way that matters to me.
5️⃣ Apple has a thousand engineers cooking hard on AI features and Tim will drop the hammer in June at WWDC.
Yousuf Jawwad retweeted
Speakers Announcement Continues...
Welcome @debuggerpk to #kcdisb
Yousuf's talk will highlight the Challenges of Distributed System Deployments, his journey of building a @qtmready Delivery Solution for Platform teams.
Conf Agenda: community.cncf.io/events/det…
Yousuf Jawwad retweeted
The xz backdoor was the final part of a campaign that spanned two years of operations. These operations were predominantly HUMINT style agent operations. There was an approach that lasted months before the Jia Tan persona was well positioned to be given a trusted role.
Yousuf Jawwad retweeted
Increasingly wondering if this AI wave is what the 90s internet wave felt like-
*Giant tech disruption
*Who wins at ~everything up in the air
*Incredibly fast changes every week in tech and capabilities
*User behavior moving fast
Everything suddenly feels up for grabs 🌪️
Yousuf Jawwad retweeted
If you received a soft commitment by a VC and the round takes many months to come together, you can’t count on those early soft commits. I’ve been part of this dynamic recently and it’s not fun for founders.
