Especially ordinary engineer
San Francisco, CA
Joined May 2010
- Tweets1.2K
- Following997
- Followers627
- Likes11.4K
πΉππππ π½ππππππ retweeted
NEW: I sat down w/ former Twitter CEO & CTO Parag Agrawal (@paraga)
Now Founder & CEO of $2B Parallel Web Systems (@p0).
Parag says we're underestimating AI agents by 5,000β10,000x, transforming the internet's economics entirely.
Forget the startup benchmark of 3x, 3x, 2x, 2x. Parag benchmarks this new agentic era against the growth of inference, roughly 10x a year, & says every estimate heβs made in this market has come in too low.
Parallel has raised $230M in total funding from funds like @IndexVentures, Kleiner Perkins & Khosla Ventures, building AI-native search infrastructure for customers like: Harvey, Lovable, Granola, Replit, LangChain, Profound, Modal, Notion & Browserbase.
Today, agent web usage is estimated at just 10β20% of human usage. Parag's own agents already run several thousand searches a day, & he believes we're still in the earliest stages of this transition.
The infrastructure implications are massive. Parag predicts that for every 10 GPUs used in model inference, another 1β2 will be required for web search alone.
The economic implications are equally significant. Traditional advertising, subscriptions & affiliate models depend on human attention and browsing behavior. As agents become the primary consumers of information, these models will need to be rebuilt, including how publishers get compensated when agents access their content.
We also covered:
βΊ Why the web is shifting from reactive to proactive AI agents
βΊ Sourcing GPUs & scaling search infrastructure
βΊ Voice agents, latency & token efficiency
βΊ Agent identity, bot detection & cybersecurity
βΊ Scaling Parallel from 30 to 120 employees this year
We also spent time on Parag's 11 years at Twitter, where he went from engineer to CTO to CEO. We discussed the lessons from leading one of the world's most influential social platforms.
ππππππππππ
(00:00) Parag Agrawal, Founder & CEO at Parallel
(01:03) What is Parallel AI?
(05:28) Bots took over the web & itβs only the start
(07:22) Will people stick with AI agents?
(11:55) Why models will get bigger, smarter & cheaper
(14:47) Where AI startups actually find GPUs
(19:43) Is 10x a year the new normal?
(21:59) Searching the web faster than you can blink
(26:02) AI that knows what you need before you do
(26:57) Agents are breaking the webβs business model
(31:08) Good bots vs bad bots
(33:45) Agents & the cyber arms race
(37:19) The webβs biggest problem with agents
(42:04) Why great models still give bad answers
(45:13) Parallel built a website for robots?
(47:56) Proactive agents & the future of search
(51:30) The person who drives Paragβs ambition
(55:50) Why feeling insecure is a superpower
πΉππππ π½ππππππ retweeted
Our team built a prototype of how Frec could seamlessly integrate with AI assistants like ChatGPT, Gemini, and Claude through the MCP protocol.
Imagine your favorite AI assistant having access to your financial picture so it can offer personalized and sophisticated wealth advice - with no competing incentives - at a tiny fraction of the cost of having a human advisor. The AI assistant may also take action by queueing up orders on your behalf, subject to your final approval in your investment app.
It's awesome to see a glimpse of the future.
Even the most disciplined investors seek some advice or second opinion about their finances. I believe that with AI, the era of paying a human advisor 1% of wealth every year is coming to an end. How can a human with limited time, memory, and context compete with LLMs and reasoning models that have access to the latest financial information, economic data, and scientific research papers? AI assistants are more likely to give better and impartial advice.
Of course, regulatory frameworks will need to adapt and model capabilities need to improve. Once that happens, investment apps which openly embrace AI will win. Our mission at Frec is to bring sophisticated investment strategies, like direct indexing, to everyone, and we want to open up to AI assistants so customers can get guidance if they choose to. We feel strongly that customers should always have the final say and approve changes queued up by AI assistants.
Every vertical consumer category will go through a transformation in the next few years. We are excited to be building the investing one.
Hereβs a video of the prototype on how to recession-proof a portfolio with the help of Claude. If youβre interested in trying it out or have thoughts to share, contact me.
We wrote a white paper on transaction costs for direct indexing, another neat read from a DS perspective. I was surprised to see just how (and how consistently) much bid-ask spreads change intra-day. frec.com/resources/blog/direβ¦
πΉππππ π½ππππππ retweeted
How does direct indexing compare to the ETF-to-ETF tax loss harvesting offered by robo-advisors like Wealthfront and Betterment? The answer: direct indexing can capture up to 2x more capital losses, resulting in significantly more tax savings.
Frec launched a direct indexing into S&P500 product yesterday, we wrote up a white paper on the approach and on fairly extensive simulation results. It's an interesting read just from a ML/data science perspective: frec.com/resources/blog/frecβ¦
πΉππππ π½ππππππ retweeted
Excited to announce @frecfinance has emerged from stealth to give everyone access to sophisticated investment products exclusively available via wealth managers.
And weβre backed by $26.4 million in funding led by Greylock and other leading investors. π§΅
frec.com
Ceiling fans are really a bad UX. The fan speed should be represented as an IEEE floating point with a pull string for each binary digit.
the youtube recommendation algorithm has learned blackmail, you WILL watch this because I will not stop recommending it until you do
πΉππππ π½ππππππ retweeted
Excited to announce that Frec is powering AngelList Treasury's money market funds integration!
You might think that kids don't learn useful skills in Roblox. But my 9yo getting scammed out of her best pet in Adopt Me actually seems like a pretty good low stakes lesson.
Really cool to have chatGPT as a second option for search queries that don't work in google. The query "how to prevent facebook from generating link previews for my website" returns nothing of value on google (missing "my website" context), but chatGPT just gives the answer.
Do videos even work on this website for anyone else? Enable sound -> video freezes, enable full screen -> video freezes, do nothing -> video is fine, then freezes.
Since this was "fixed" my first 5-10 tweets on my timeline are always a tweet by Elon or a "Elon replied" injection, whether they are new tweets or a 1-2 days ago.
Long day at Twitter HQ with eng team
Two significant problems mostly addressed:
1. Fanout service for Following feed was getting overloaded when I tweeted, resulting in up to 95% of my tweets not getting delivered at all. Following is now pulling from search (aka Earlybird). When Fanout crashed, it would also destroy anyone elseβs tweets in queue.
2. Recommendation algorithm was using absolute block count, rather than percentile block count, causing accounts with many followers to be dumped, even if blocks were only 0.1% of followers. Also, itβs trivial to bot spam accounts with blocks.
