Bryan Bischof fka Dr. Donut (@BEBischof) header
Bryan Bischof fka Dr. Donut (@BEBischof) profile picture

Leading AI @theoryvc; Teach ML @rutgersu. Prev: Led AI @_hex_tech; Led Data @weights_biases, ML @stitchfix, Led Data @bluebottleroast

Ohlone Territory
Joined December 2010
Photo by Bryan Bischof fka Dr. Donut (@BEBischof)Photo by Bryan Bischof fka Dr. Donut (@BEBischof)Photo by Bryan Bischof fka Dr. Donut (@BEBischof)Photo by Bryan Bischof fka Dr. Donut (@BEBischof)Photo by Bryan Bischof fka Dr. Donut (@BEBischof)Photo by Bryan Bischof fka Dr. Donut (@BEBischof)Photo by Bryan Bischof fka Dr. Donut (@BEBischof)Photo by Bryan Bischof fka Dr. Donut (@BEBischof)Photo by Bryan Bischof fka Dr. Donut (@BEBischof)Photo by Bryan Bischof fka Dr. Donut (@BEBischof)
A rat can run a restaurant, but can an agent run a bakery? Welcome to The Great Business Bake Off: our AI Agent Hackathon. On December 11, you and your agents get five hours to make the most profitable bakery in town. Spots are limited: theoryvc.com/bakeoff
43
32
32
148
44,855
Bryan Bischof fka Dr. Donut retweeted
this. so much. invest in your next gen data architecture. use sql or lance to join instead using agents to join a dozen different data sources.
a context layer is actually not counter to the bitter lesson it’s just systems infra: a local read replica + derived indexes built by offline LLMs (no hardcoded ontologies, just recompute as models get better) people love to cite the bitter lesson as “skip the infra, let the model figure it out live,” but search is literally half of sutton’s point CPUs getting orders of magnitude faster didn’t kill b-trees or materialized views. those exist because compute outpaces I/O let's assume frontier-level-intelligence inference hits 0ms: live fan-out across scattered SaaS APIs still leaves you I/O-bound. and then in the SaaS endgame, proprietary state is the last moat they defend so those API straws only get narrower. narrow pipes still work with slow incremental sync of raw state, but not parallel search at runtime everyone fanning out is already eating 429s and it'll probably only get worse if you keep raw ground truth synced underneath so runtime models can search and verify in milliseconds + pay compute at write-time and read-time --> hooray you can still be bitter-lesson pilled with a context layer
2
4
24
1,290
Bryan Bischof fka Dr. Donut retweeted
It’s important to take breaks from video games; every 30minutes read 5 pages of Evals 😅
I got about 8% of the way through GTA (too distracted with side quests) and 30% of the way through EFAIE.
10
3
36
3,214
Bryan Bischof fka Dr. Donut retweeted
person who thinks AI will prove the Riemann hypothesis in 2026 but that Apple won’t deliver an LLM-powered Siri until 2027
183
523
198
6,420
564,974
Bryan Bischof fka Dr. Donut retweeted
OK, got some updated results from the team on e-means. We implemented hierarchical e-means and now it handily beats our existing hierarchical k-means. The speed up diminishes by a lot at high dimensionality (128 vs 1024), which unfortunately is where most embedding models are at, but i think it's still substantial enough. @swann_bessa do these results pass the smell test? (ignore the main vs branch thing, that's just some incremental optimization we added separately on kmeans)
Interesting new work on clustering algorithms. We did a quick benchmark and preliminary results show that hierarchical kmeans (Lance default) is still much faster. Curious if anyone else has done similar comparisons already.
3
3
20
1,104
Bryan Bischof fka Dr. Donut retweeted
Btw Alessandro Masullo proposed a proof of GRH several months ago that has been under review by Brian Conrad "... and it is so far looking good." math216.wordpress.com/2026/1…
50
144
71
1,301
280,128
Bryan Bischof fka Dr. Donut retweeted
1
6
295
Things are heating up in the kitchen!
Some knights guard castles. Ours guards the bakery. Join us in SF on December 11 to build a bakery with an agent alongside engineers and data teams. Space is limited: na2.hubs.ly/H0955TB0 @theoryvc
1
5
358
Bryan Bischof fka Dr. Donut retweeted
So close to GRH before GTAVI.
7
67
1,867
Claim of Riemann Hypothesis by alpoge alpo.ge/remark.pdf
3
4
3
23
6,114
Bryan Bischof fka Dr. Donut retweeted
This mom draws Pokemon using only their online descriptions (via pok.emum/IG)
391
6,511
638
93,686
6,082,470
Bryan Bischof fka Dr. Donut retweeted
Here is a slide deck I once made about my name, listing many amusing encounters I've had with people of the same name docs.google.com/presentation…
In second or third or fourth grade, a man named Tao Lin came in to speak to our class. I'd never met or heard of another Tao Lin. I felt embarrassed. Today, I see there are hundreds, maybe thousands, of Tao Lins, and I no longer feel embarrassed by the existence of other Tao Lins.
1
12
976
ngl this has potential to be a really great approach to viz.
1
16
1,617
Since many people have asked... the answer is yes:
🤖 Made with AI
2
1
18
645
“When you take into consideration all the expensive maintenance, is it really worth buying an older car?” Over 8 years ago I bought my 2003 GX470 – what was essentially my dream car. Already 15 years old with 167k miles on it and a few known issues. Now I’m ready to move on from it, but I wanted to see how the total cost of ownership really broke down; where did every dollar go? I actually have that data, so here’s the chart.
12
23
1,871
Bryan Bischof fka Dr. Donut retweeted
Just noticed "recursive self-improvement" contains "rsi" in the middle of it...
69
118
15
2,724
92,184
Bryan Bischof fka Dr. Donut retweeted
I got another sneak peek behind this Hackathon and its going to be an insane and fun challenge. There will be evals! In addition to judging, @sh_reya and I are giving away signed copies of our AI Evals book at the event Sign up here: luma.com/bakeoff?utm_source=…
What happens when software engineers and data teams meet in San Francisco to build their own bakery with an agent? The answer is.. we'll ducking see. Space is limited, join us in December if you're around: luma.com/bakeoff
19
3
46
4,753
Bryan Bischof fka Dr. Donut retweeted
Last night I was awoken by my agent. I asked it to reverse engineer airplay over night it decided at 3 AM that the right way to verify that it works was to listen on the laptop mic to a sound played at max volume in the living room. It raised the volume to hear it.
136
104
34
5,122
216,566
Me and @xeophon were just talking about making this meme. All my best ideas get scooped
Benchmark naming is nearly saturated!
3
339
Bryan Bischof fka Dr. Donut retweeted
friends, i've been thinking a lot about the great sidebar design challenge of our time and have come to bring you the fruits of my research!
19
64
6
1,166
54,367