Tomasz Tunguz
@ttunguz
Venture capitalist at @theoryvc
Student of Startups
Backer of 9 unicorns
Author of
Subscribe
362 Following 129.5K Followers
GPU prices doubled from $4 to $8 per hour. OpenAI cut Luna's prices by 80%, then another 50%.
We're getting more work out of each GPU hour, enough to keep prices falling.
How long will this last?
Show more
AI models are improving faster than we can understand them.
Dario Amodei wants to narrow that gap by pacing development, but the consequences extend well beyond the models.
We’re starting a new series called Feature of the Week 🎉
This week’s feature by @Yah01_: Index prewarm now reads in parallel bytes windows. Loading a billion-vector index into memory used to take 95 minutes, now it takes 3.
Show more
The better the harness, the better the business.
Berkeley published a study this week showing harnesses, the systems that control AI agents, set the price of an answer. The right harness cuts the cost of the same result by 71% without a loss of accuracy.
Show more
Huge milestone for @seanmcccarthy & the @Backopsai team! Congratulations on the Series B 📦
Moving a shipment takes more than a truck.
Pickups need confirming, documents need checking, and changes need coordinating.
We sat down with @Backopsai CEO @seanmcccarthy to talk about why logistics is still so manual and what his team is doing to change that.
Show more
Vercel took its inbound sales development team from 10 people to 1.25, increased productivity by 30% & it costs just a few thousand per year in tokens.
Dario Amodei asked the industry to pace itself on a Saturday. But what does that word mean?
Pacing is a policy question. Five camps priced the consequences & none named a speed.
Show more
DuckDB 2.0 alpha reads Parquet and CSV from S3 2x to 3x faster, with zero query changes.
1.5: each worker downloads, waits, decodes, repeats.
2.0: a download pool fetches ahead, workers only decode.
Same laptop, same query: 18.8 s to 7.7 s. On by default.
Full numbers, plus recursive CTEs and VARIANT:
Show more
ChatGPT Desktop (the Codex app) can now be configured to use Ollama models. Download or update to Ollama 0.34 to get started.
AI consumption is not a smooth curve anyone can extrapolate : it arrives in three waves, each an order of magnitude larger than the last.
The second one has already broken.
The third is on the horizon.
Show more
We nearly doubled MFU on TPU v6e serving Gemma 4 31B, taking prefill from 32% to 63% MFU, and 2x throughput.
v6e has 2.5x less HBM and less than half the memory bandwidth of an H100.
@transmissions11 wrote it up here:
Show more
DATA OUTPOST: A 2-day conference for data people in a world being rebuilt by AI.
If you are a founder, engineer, product leader, researcher, operator, or data architect making machine intelligence useful inside actual companies, consider this your forward station.
Nov 4: Hands-on-keyboard workshops. Less talking about building, more building.
Nov 5: Talks and panels including:
Tristan Handy (Fivetran + dbt Labs)
Stefanie Tignor, PhD (Clay)
Sudeep Dasgupta (DoorDash)
Tosh Rayadhurgam (Stripe)
Jordan Tigani (our own!)
...and more.
Tickets are $499 for both days:
Show more
Can't wait to chat with @jdewitt29 !!!
@jdewitt29, COO of @vercel, joins @ttunguz to walk through something that kept failing a few years ago: running inbound sales with AI.
📅 Oct 1
⏰ 11am PT
📝 Register here:
Show more
The market is telling us that we should be 3x more productive with AI.
What if that productivity gain is just an AI working 24 hours a day while a human works eight?
Over the next five years, US data center capacity will grow from 25 gigawatts to 70 gigawatts, part of a global buildout costing roughly $5t.
Where will the money come from?
Model companies buy electricity by the megawatt and resell it as cognitive work.
Anthropic's revenue has run as high as $50m per megawatt, against $10-15m of compute.
Meta launched two things yesterday : a state-of-the-art model & a new pricing system for foundation models. Muse Spark propels US open source models to the frontier.
Meanwhile, the pricing system resets the industry’s economics.
Show more
Enormously important for US to have SoTA open models
Muse Spark 1.3 is rolling out today with frontier performance almost too cheap to meter. This is the biggest jump we've made so far on coding and agentic work. Try it in Muse Code and our API.
Next up 🍉 and Muse Spark open weights releases coming soon.
Show more
The frontier AI market is sorting itself into closed camps, as labs pick partners, cut rivals, & ration access to their strongest models.
OpenAI's new Luna pricing now even below DeepSeek on intelligence/dollar.
