bad cop @autoresearchinc
San Francisco, CA
Joined December 2019
- Tweets95
- Following117
- Followers152
- Likes659
like they still do a lot of stupid things, but i'm finding myself thinking "huh, this is better than what i would've done" quite often when reading llm generated code these days.
the cycle of new benchmark released -> labs train on related tasks -> benchmark saturated -> new benchmark created has been going for a while now. it seems inevitable that anything benchmarkable will topple. here's how to leverage this cycle to get rich quick:
1. pick a domain that isn't saturated yet, create a benchmark for it
2. build RL environments + tasks for that benchmark. sell them to the big labs, use the $ to fund product development
3. build a product that hinges on LLMs solving that problem well
4. next model release saturates it, product works, you have a head start
Thomas Boser retweeted
We benchmarked the vulnerability-finding capabilities of Codex Security, Claude Security, and Nebu across coverage, cost, and time.
Our results:
Nebu (GPT-5.6-Sol) > Nebu (GPT-5.4) ≥ Codex Security (GPT-5.6-Sol) > Claude Security (Opus 5) > Codex Security (GPT-5.4)
A strong security pipeline amplified model performance severalfold, allowing an older GPT-5.4 model to match Codex Security running GPT-5.6-Sol, while completing the analysis faster and at a lower cost.
See the full report in comment.
Thomas Boser retweeted
Today we’re announcing r-1, our new document parsing model. It’s more accurate than our most powerful agentic OCR models, faster, and up to 6x cheaper.
At @reductoai, we spent two years building specialized models for complex visual layouts, tables spanning multiple pages, and key formatting like strikethroughs. We then used everything we learned to build r-1, the first in a new generation of models designed to handle the hardest documents without multiplying cost.
This early preview delivers a 20% lower error rate than our most accurate legacy agentic models and will keep improving with new checkpoints over the next few weeks.
Accuracy is only the beginning. r-1 preview is available at 1¢ per page all-in, with additional volume discounts as you scale. In the near term we’re also going to release r-1 mini, and an auto mode that intelligently selects the right approach for each page.
If you’re currently using another parser, we’re offering up to $5,000 in credits to evaluate and migrate. You can claim the migration offer and learn more about r-1 using the links in the comments.
Happy parsing!
Thomas Boser retweeted
their margin is our opportunity
lowermyanthropicbill.com
‘Investors’ think Anthropic can reach a $120B revenue run rate by year end
Six of its investors told the FT this, and honestly, I can see the path.
Reaching $120B would mean hitting a $10B monthly revenue pace by December, or growing another 2.55x from May. Remember Anthropic grew more than 5x in the five months before May.
One investor thinks even a “30x revenue multiple” would be on the low end and would make Anthropic a $3T company.
The obvious caveat is that these are six unnamed investors using their own models, not Ant management running their own financial models for its planned September/October IPO. And $120B means annualized revenue, not revenue actually booked during 2026.
I would take it with a grain of salt I think that a much more likely to hit 80B ARR which is still crazy.
Dinner or not, I’d love to chat with folks interesting autoresearch at AI4 next week!
What happens when AI systems can improve themselves?
At Ai4, we’re bringing together founders and AI/ML leaders for a small dinner on systems that learn from real-world use, run their own experiments, and keep getting more accurate, faster, and cheaper.
Join us: luma.com/q7u2yg9w
Thomas Boser retweeted
Every AI company should own its data, its models, and the research lab that keeps making them better.
Today, @thomasboser and I are launching @hiloopai to make that possible.
The frontier advantage isn’t access to a model. It’s the research organization continuously improving it.
Your product already generates the raw material for better intelligence: proprietary data, production feedback, evaluations, and domain expertise. Very few teams have the research capacity to turn those assets into better models.
Hiloop builds and operates that capability with you.
Bring us the model your product depends on, the data that makes it different, and an evaluation that defines success. We reproduce your baseline and run an autonomous research campaign against it.
Agents pursue competing hypotheses in parallel, build on previous results, and promote only improvements that survive verification. Our researchers validate the winners and bring them into production.
The work can run hosted or inside your environment. You retain control of your data, and the resulting models, evaluations, and research artifacts are yours.
We’re starting with model training, post-training, and inference optimization, where progress is measurable and the value is immediate.
But this isn’t a one-off model improvement. It’s a persistent research capability that begins each campaign with everything learned from the last.
Over time, the lab accumulates research memory and improves its own tools, evaluations, and agents. The process used to create better intelligence gets better itself.
That’s infrastructure for recursive self-improvement.
If your product depends on a model and an important metric has stopped moving, bring us the model you can’t make better. We're much better at research than making videos!
Reach out to us: [email protected]
A more detailed overview of what we're doing at hiloop. Something we don't really mention - we're also making this secure for enterprise customers (data exfiltration controls, fully deployable in your cloud, etc. Reach out to learn more!
Check out our launch YC post for more details on what we're doing!
ycombinator.com/launches/SBN…
We're still working on our cinematography skills, but we're pretty good at autoresearch. Hit us up if that's something you want to get better at too!
Every AI company should own its data, its models, and the research lab that keeps making them better.
Today, @tboser and I are launching @hiloopai to make that possible.
The frontier advantage isn’t access to a model. It’s the research organization continuously improving it.
Your product already generates the raw material for better intelligence: proprietary data, production feedback, evaluations, and domain expertise. Very few teams have the research capacity to turn those assets into better models.
Hiloop builds and operates that capability with you.
Bring us the model your product depends on, the data that makes it different, and an evaluation that defines success. We reproduce your baseline and run an autonomous research campaign against it.
Agents pursue competing hypotheses in parallel, build on previous results, and promote only improvements that survive verification. Our researchers validate the winners and bring them into production.
The work can run hosted or inside your environment. You retain control of your data, and the resulting models, evaluations, and research artifacts are yours.
We’re starting with model training, post-training, and inference optimization, where progress is measurable and the value is immediate.
But this isn’t a one-off model improvement. It’s a persistent research capability that begins each campaign with everything learned from the last.
Over time, the lab accumulates research memory and improves its own tools, evaluations, and agents. The process used to create better intelligence gets better itself.
That’s infrastructure for recursive self-improvement.
If your product depends on a model and an important metric has stopped moving, bring us the model you can’t make better. We're much better at research than making videos!
Reach out to us: [email protected]
There's a new version of this post
Thomas Boser retweeted
A bit of a late update, but @thomasboser and I got into YC S26. We're building @hiloopai, and we got a pretty cool result. Check it out!
@deepmatmul pointed agents at Karpathy's autoresearch benchmark and achieved a SOTA result.
We're building infrastructure to scale autoresearch. We work with teams on their hardest problems. Hosted or on-prem. Reach out to us: [email protected]
Life update - I’m now working on building infrastructure for scaling autoresearch.
@deepmatmul has been hard at work and has already achieved SOTA results on the original autoresearch task!
If you’re solving hard problems we’d love to work with you!
