
Product, AI, startups, retail. CPO @bookendai. Led Retail Supply AI @googlecloud. Cofounded @incentivetarget (acq, @Google). Led product @catalina, @HealthEdge.
Boston, MA
Joined April 2009
- Tweets11.6K
- Following300
- Followers839
- Likes15.6K
Ben Sprecher retweeted
NEW ⚡: We let Opus 5.5 run an AI company for seven weeks.
We gave it an HF account (Horizon-Labs) and compute, and asked it to earn real adoption by creating useful models.
It reached 50k downloads and its models are being used in more than 100 GitHub projects.
Amazing
In my ongoing quest to find the limits of Opus 5.5, I directed it to audit my old scientific field, theoretical astrophysics. We indexed about 150,000 papers, audited 50,000 for issues, and did progressively more thorough checks on thousands more. We found nearly 1000 major errors that got cited many times, most are still uncorrected. This survey enabled me to estimate the efficacy of peer review for correcting errors (not very good) and to identify dozens of open questions that anyone with spare Claude Max tokens could answer today (.md context files included).
Bottom line: AI can speed up quantitative science by >1000x with no further model improvement required. But I'm seeing no evidence of this being pushed within universities, only academics complaining about labs dropping math proofs.
Among my fellow reformed academic high agency founders, however, several are speed running entire disciplines.
This survey has implications for how science will be run in the very near future. Hint: Academic publishing is dead. It was already a very stupid way to record and promulgate knowledge, and had been so for decades, even before all that publicly funded knowledge was locked behind paywall cartels. Now AIs can generate thousands of novel, correct papers per day, it is no longer meaningful for humans to consider a research paper a "unit of science" or to even bother much with the 99.9% of science which, up to this point, was necessary drudgery. Time to discover something like General Relativity or Evolution every week. LFG!!
Enjoy: science.caseyhandmer.com/phy…
34
Ben Sprecher retweeted
In my ongoing quest to find the limits of Opus 5.5, I directed it to audit my old scientific field, theoretical astrophysics. We indexed about 150,000 papers, audited 50,000 for issues, and did progressively more thorough checks on thousands more. We found nearly 1000 major errors that got cited many times, most are still uncorrected. This survey enabled me to estimate the efficacy of peer review for correcting errors (not very good) and to identify dozens of open questions that anyone with spare Claude Max tokens could answer today (.md context files included).
Bottom line: AI can speed up quantitative science by >1000x with no further model improvement required. But I'm seeing no evidence of this being pushed within universities, only academics complaining about labs dropping math proofs.
Among my fellow reformed academic high agency founders, however, several are speed running entire disciplines.
This survey has implications for how science will be run in the very near future. Hint: Academic publishing is dead. It was already a very stupid way to record and promulgate knowledge, and had been so for decades, even before all that publicly funded knowledge was locked behind paywall cartels. Now AIs can generate thousands of novel, correct papers per day, it is no longer meaningful for humans to consider a research paper a "unit of science" or to even bother much with the 99.9% of science which, up to this point, was necessary drudgery. Time to discover something like General Relativity or Evolution every week. LFG!!
Enjoy: science.caseyhandmer.com/phy…
Ben Sprecher retweeted
Fascinating data from Stripe on the "SaaSpocalypse."
"SaaS equity prices have since returned to roughly pre-SaaSpocalypse levels."
And: "recent SaaS revenue growth has remained strong and even accelerated."
This echoes what we're seeing across @tinyseedfund’s 217 SaaS portfolio companies.
It doesn't mean AI won't reshape SaaS. It will. No one knows how much. And times of massive change throw off confusing signals like these.
Growth is strong for many, yet some companies have been devastated. M&A multiples are down because no one can price long-term viability.
All of these things are true.
Link to the Stripe article below👇
Ben Sprecher retweeted
Concurrent Agents / GPU
I used to wonder if the whole agent thing was an anthropomorphism of something that actually wants to be a commodity.
I thought about this quite a bit, because I had to solve an AI problem back in 2016 and whilst I wasn’t using LLMs and didn’t have anything like agents the solution that worked was to break by problem in 10-20 sub problems, splits my training data into subdivisions and train specialists models for each subdivisions. I then ran a classifier layer over the top and would allocate the specific model to solve the specific subset of the problem… this worked.
Agents are an LLM version of something a bit more generalised in problemspace, which is “how to eat an elephant”.
I think LLMs are close to the .py point where LLMs and agents just become the status quo because it’s the AI branch that has the most developers.
I think it gets harder and less likely for an alternative technology to steal the frontier the longer LLMs spend being adapted and integrated into the economy. I think those Hail Mary models have an ever larger hill to climb for adoption.
So whilst agents are very 2026-coded, they’re probably going to stay.
Now…
Tokens are a crappy metric and a poor choice for monetisation. Forgivable, as it was all we had back then. But nobody knows what a token is worth and no two tokens have the same value, tokens price infrastructure not models.
Agents are where I think the monetisation business models will move to in 2027, and this makes a lot more sense to any business.
You can price an agent by the competence of the model it operates, if you sell your users something like “maximum concurrent agents” you can sell over 100% of your capacity. This is nice when you are capacity constrained as will be the case for a long time.
What is interesting for the GPU is that concurrent agents are not linear with tokens/s
Value is some function of:
• Model capability (which is non-linear)
• Agent concurrency
• Token bandwidth
I increasingly think this is a hardware optimisation problem.
Is your hardware optimised to serve a faster stream of tokens, or is it optimised to serve more concurrent agents? Sounds the same, but no it is not the same.
I think compute will be increasingly heterogeneous and agents themselves will also become increasingly heterogeneous.
I often wonder if this is proximity bias from stuff I was doing in 2016-2020, but I don’t think so. The bitter lesson suggests bigger plastic lump just wins, but I think once you apply cost realities and profit motive you have to be more sophisticated.
The market is not a performance maxxer, it is a resource allocator. If you can get $100 of value for a $50 cost, or $110 of value for $90 cost… the market will select the former.
Heterogeneity is efficient resource allocation by another name.
I also strongly believe that the most powerful and efficient use comes from deeply heterogeneous use of models. Smaller cheaper models are faster and more concurrent. Parallelise where you can.
Divide and allocate.
It sometimes feels a bit arrogant to presume that things will develop based on how I use the technology day to day, but I think this.
For transparency… my API calls cover a broad spectrum of model capabilities. You want a classifier that runs daily tests against API end points to determine performance/$.
Final thought, I don’t think the singularity will be measured by soaring intelligence going vertical.
The singularity is the point where the economy uncouples from demographics.
This is going to happen in America in the next few years, GDP is going to accelerate and break the 5,000 year link to demographics. Because demographics will no longer be limiting.
Bond market is unprepared for this, they should take this much more seriously.
Thinking about concurrent agents instead of tokens/s is one step closer to that discussion.
Ben Sprecher retweeted
the key insight i learned from @mgarimella was to design goals around what you can give feedback on more easily
feedback (used to align the agent with your mental model of what it should do) is the hardest thing to extract from users
so the more you can design your system (ui, goals, etc) around making that easy to do - the better!
as agents move towards working on goals, it becomes important to design those goals correctly
wrote with @mgarimella on some lessons learned at actively!
As I struggle to remember high school chem to help my daughter in hers, I wish they had taught us this.
Why does nuclear energy exist?
Let me explain… first of all the periodic table is not a great way to explain what elements are. A much better illustration is a ‘nuclide chart’.
In a nuclide chart we have:
protons on X axis
neutrons on Y axis
We have tiny atom elements like hydrogen and lithium at the bottom left and big atom elements like uranium and plutonium at the upper right.
Now what is interesting is that for some reason (strong force, subatomic particles and quantum physics)… the tiny atoms that are stable have a neutron:proton ratio of approx 1:1 and the big jumbo atoms like U and Pu have a neutron:proton ratio of about 1.5:1
This means the bug heavy atoms are neutron rich compared to the smaller atoms.
You can see this in the curvature of the stability line as it bends away from the 1:1 ratio.
Why does this matter?
Well it means that whenever a big atom decays (as the majority of atoms do), then it decays into 2 smaller atoms, but as it does these two smaller atoms have a lower neutron:proton ratio for the two product atoms to be stable.
This means that there are some left over neutrons that cannot be part of the two product atoms…
As a result these spare neutrons are released, and because of their tiny size and tiny mass they whizz off at high speed.
These fast moving spare neutrons then crash into other heavy atoms and split them apart too.
This isn’t spontaneous decay but is induced fission.
120 years ago, people realised this is a runaway chain reaction and that sparked the dawn of the atomic age and the geological marker of the holocene.
Now you understand why/how it happens.
4
217
Hypothesis:
- Cloud 1.0: Flight to Convenience (s3, ec2)
- Cloud 2.0: Flight to Capability (BigQuery, Vertex AI)
- Cloud 3.0: Flight to Safety/Quality (AI-driven cyber threats)
Only the hyperscalers will have the resources, talent, and AI defenses to keep up.
1
57
Ben Sprecher retweeted
Request for product
Waymo or uber feature where I can turn off prices in the app + charge it to my credit card so that my very frugal elderly parents will use it without freaking out at the prices
Ben Sprecher retweeted
Introducing Gemini Decision 1, our new frontier decision model in the Gemini API.
It is $0.10 per 1M input tokens with no additional cost. With support for image, audio, video, and PDF natively, 256K context window, & 1K questions.
You can try it for free in the API!
Pretty amazing results
Introducing e-means, an algorithm beating k-means in 10x fewer epochs.
With order of magnitude wall-clock speedups, e-means is released in FAISS and needs no hyperparameter tuning beyond k. It is robust to various scales and datasets. Still skeptical ? Just try e-means [1/10]🧵
1
88
This is kind of a big deal
Formal Verification of Signal’s Protocol by @GoogleDeepMind share.google/CXCrpJlIkEGmbNx…
30
Wtaf
my vocal synth models the vocal tract as a network of branching acoustic tubes, unlike vocaloid which stitches together and pitch-shifts phoneme recordings. still working on improving naturalness by tuning it against recordings of real singers, but I'm already impressed at how cool it sounds!
1
53
Ben Sprecher retweeted
my vocal synth models the vocal tract as a network of branching acoustic tubes, unlike vocaloid which stitches together and pitch-shifts phoneme recordings. still working on improving naturalness by tuning it against recordings of real singers, but I'm already impressed at how cool it sounds!































