Tomasz Tunguz
1,755 FOLLOWERS
Tomasz Tunguz is a venture capitalist at Redpoint since 2008. He was a PM on the Ads team at Google and worked at Appian before. He loves to learn about how businesses are created and aim to share those learnings here.
Tomasz Tunguz
2d ago
Should software assist humans or act on their behalf?
In 2016, the question was easy to answer : sell Ironman not Robocop. Technology hadn’t reached the level of sophistication we have attained today where AI is 90% as capable as a high-school student, the MMLU benchmark for AI is precisely this.
The next generation of software startups have a strategic question with different terminology & potentially a different conclusion.
To be or not to be an agent, acting on behalf of workers?
Copilots, like Github’s, complete their humans’ sentences in code, an AI pair programmer. Copilots have prov ..read more
Tomasz Tunguz
1w ago
Will AI sofware companies operate with better or worse profitability than a classic SaaS company?
Initially, I thought worse since the expense of serving AI as a product is signficantly higher.
But now I’m not so sure. AI SaaS may be much more profitable than the -10% average net income margins of the current crop of public businesses.
Yes, AI inflates the cost to serve the product. Google queries may be 10x more expensive than standard search results. That’s an unfair comparison since Google has focused on classic query cost optimization for more than 20 years.
But let’s disregard that.
AI is ..read more
Tomasz Tunguz
1w ago
Within data teams, a tension exists. Centralize the data analysis to ensure accuracy or enable end-users to analyze their own data directly which is faster & more direct.
The pendulum between these two states started with centralization during the 2000s with BI products from Microstrategy, Cognos, BusinessObjects, & Hyperion. In 2004, Tableau emerged from the Stanford campus to deliver their application to the users.
Cloud databases ushered in an opportunity to centralize that data analysis again. Looker’s modeling language, LookML, provided a way to define metrics across an organizat ..read more
Tomasz Tunguz
1w ago
If you were to watch three videos on YouTube Shorts - one on Italian cooking, one on chess openings, & a third on crypto trading, YouTube Shorts’ recommendation algorithm combines the video descriptions with your dwell time.
Watching the osso bucco video to its end would trigger more Italian cooking specialty videos in your feed.
We believe every LLM-based application will need this capability.
Combining text & structured data in an LLM workflow the right way is difficult. It requires a new software infrastructure layer: a vector computer.
Vector computers simplify many kinds of data ..read more
Tomasz Tunguz
2w ago
The database is being unbundled. Historically, a database like Snowflake sold both data storage & a query engine (& the computing power to execute the query). That’s step 1 above.
But, customers are pushing for a deeper separation of compute & storage. The recent Snowflake earnings call highlighted the trend. Larger customers prefer open formats for interoperability (step 2 & 3).
A lot of big customers want to have open file formats to give them the options…So data interoperability is very much a thing and our AI products can generally act on data that is sitting in cloud sto ..read more
Tomasz Tunguz
2w ago
For those of us who love logic, the paradoxical title of this post should catch your eye, just as it did mine.
In Alchemy, the founder of a major brand agency describes the way many of the major consumer companies in the world created brands.
We call it breaking out : a double entendre which means both growing faster than competitors but also in a different way than their competition.
Rory Sutherland, the author explains it this way :
“The fatal issue is that logic always gets you to exactly the same place as your competitors.”
So he advocates to
“Test counterintuitive things, because no on ..read more
Tomasz Tunguz
3w ago
On March 15th at 9:30 Pacific time, Office Hours will host Colin Zima, CEO of Omni Analytics.
Colin is no stranger to business intelligence & data analysis. He worked on search quality at Google, founded a dynamic pricing company for the restaurant industry, then ran data at a HotelTonight before becoming Chief Analytics Officer at Looker through its acquisition by Google.
Colin has advised many of the world’s largest companies on their BI strategy.
Since Looker’s acquisition, the business intelligence market has evolved with the advent of new databases, new hybrid technical architectures ..read more
Tomasz Tunguz
3w ago
In the past, the bigger the AI model, the better the performance. Across OpenAI’s models for example, parameters have grown by 1000x+ & performance has nearly tripled.
OpenAI Model
Release Date
Parameters, B
MMLU
GPT2
2/14/19
1.5
0.324
GPT3
6/11/20
175
0.539
GPT3.5
3/15/22
175
0.7
GPT4
3/14/23
1760
0.864
But model performance will soon asymptote - at least on this metric.
This is a chart of many recent AI models’ performance according to a broadly accepted benchmark called MMLU. 1 MMLU measures the performance of an AI model compared to a high school student.
I’ve categori ..read more
Tomasz Tunguz
3w ago
Last week, Reddit filed their S-1 to go public. At least 10% of their revenue - about $60m - comes from selling data to train Large Language Models. Reddit’s data sales revenue will likely be much more than 10% by the end of the year. Quoting directly :
We expect our growing data advantage and intellectual property to continue to be a key element in the training of future LLMs.
This raises a fundamental question : What if the revenue from data sales dwarfs the revenue from ads?
LLMs need data. They compress this it & reconstitute it to answer user queries. At Reddit, like many sites on t ..read more