Pinned Tweet
Proximal builds infrastructure that enables AI to improve from real-world experience.
In the past year, we have grown to more than $200M in annualized revenue helping frontier labs and enterprises improve coding agents. Now, we are expanding beyond software engineering.
Recent demonstrates what happens when models are trained in domains with massive amounts of public data and easy verifiability, making it easy to find weaknesses, generate training tasks, and iterate quickly.
We build infrastructure that enables this feedback loop in other domains: we are excited about a world in which AI systems design targeted drugs, accelerate chip design, and rewrite legacy software that still runs hospitals, power grids, governments, and other critical infrastructure.
Since starting, we have assembled a small team of researchers and engineers and built world-class infrastructure for post-training and synthetic data research to overcome the challenges of scaling manual data curation.
Our team comes from Cursor, Google DeepMind, Meta Superintelligence, Prime Intellect, Citadel, and Jane Street. More than half are former founders.
We’re fortunate to be backed by investors who share that vision, including @generalcatalyst who led our $15M Seed Round at a $300M valuation, as well as @chemistry, @svangel, @dvlamoen and individuals like @LiamFedus, @kevinweil, and @bernhardsson
In the past year, we have grown to more than $200M in annualized revenue helping frontier labs and enterprises improve coding agents. Now, we are expanding beyond software engineering.
Recent demonstrates what happens when models are trained in domains with massive amounts of public data and easy verifiability, making it easy to find weaknesses, generate training tasks, and iterate quickly.
We build infrastructure that enables this feedback loop in other domains: we are excited about a world in which AI systems design targeted drugs, accelerate chip design, and rewrite legacy software that still runs hospitals, power grids, governments, and other critical infrastructure.
Since starting, we have assembled a small team of researchers and engineers and built world-class infrastructure for post-training and synthetic data research to overcome the challenges of scaling manual data curation.
Our team comes from Cursor, Google DeepMind, Meta Superintelligence, Prime Intellect, Citadel, and Jane Street. More than half are former founders.
We’re fortunate to be backed by investors who share that vision, including @generalcatalyst who led our $15M Seed Round at a $300M valuation, as well as @chemistry, @svangel, @dvlamoen and individuals like @LiamFedus, @kevinweil, and @bernhardsson
113
42
388
260,991
Gemini 4 Carbon scores 61.2% on FrontierSWE
The model is only outperformed by GPT-6 Astra (65.5), Opus 5.5 (62.3) and Sonnet 5.5 (61.9)
The model is only outperformed by GPT-6 Astra (65.5), Opus 5.5 (62.3) and Sonnet 5.5 (61.9)
9
7
121
12,200
Proximal is excited to be an official partner to @baseten.
We're looking forward to leveraging our research and expertise on data, evals and post-training to help enterprises own their intelligence.
We're looking forward to leveraging our research and expertise on data, evals and post-training to help enterprises own their intelligence.
3
7
84
12,073
In September, we published the TORCS Racing Bot task as part of FrontierSWE
Agents get 20h to build a bot that drives in a racing simulator as fast as possible, using only visual inputs. We are sharing a more detailed analysis and interesting findings from models' attempts
Agents get 20h to build a bot that drives in a racing simulator as fast as possible, using only visual inputs. We are sharing a more detailed analysis and interesting findings from models' attempts
11
9
141
17,622
Overall, rule-based methods were still competitive with neural approaches. No agent was able to come close to our human baselines, indicating room for improvement
1
0
10
528
Proximal retweeted
We are hiring for the following roles (DMs open!):
1) Research Engineer: Run eval and training experiments to understand how to improve model capabilities
2) Platform Engineer: Build infrastructure to orchestrate millions of agents and sandboxes per day
3) Technical Generalist: Work closely with our enterprise and frontier lab customers and own projects end to end
1) Research Engineer: Run eval and training experiments to understand how to improve model capabilities
2) Platform Engineer: Build infrastructure to orchestrate millions of agents and sandboxes per day
3) Technical Generalist: Work closely with our enterprise and frontier lab customers and own projects end to end
At Proximal, we believe that building our own research infrastructure is crucial to design the highest-quality training data
Today, we are sharing more information about our internal post-training stack
26
26
469
25,748
At Proximal, we believe that building our own research infrastructure is crucial to design the highest-quality training data
Today, we are sharing more information about our internal post-training stack
Today, we are sharing more information about our internal post-training stack
32
43
798
164,601
Using this setup, we trained Qwen3.8-27B on an internal set of coding tasks.
The model's Pass@1 on DeepSWE climbed to 37.2%, 8.4 percentage points above the base model. Pass@8 increased from 67.7% to 86.6%.
The model's Pass@1 on DeepSWE climbed to 37.2%, 8.4 percentage points above the base model. Pass@8 increased from 67.7% to 86.6%.
2
1
25
4,491
If you want to help us build world-class infrastructure, please reach out. You can read the full blog post here:
proximal.so/blog/posttrain…
proximal.so/blog/posttrain…
0
3
42
4,243
We have added Claude Haiku 5.5 and Sonnet 5.5 to the FrontierSWE leaderboard.
Both models are extremely capable despite being smaller than Anthropic's flagship models Opus and Fable - Haiku outperforms GPT-5.6 Sol, GLM-5.3 and other strong models!
Both models are extremely capable despite being smaller than Anthropic's flagship models Opus and Fable - Haiku outperforms GPT-5.6 Sol, GLM-5.3 and other strong models!
14
3
95
12,084
Proximal retweeted
If you are interested in:
- Post-Training
- Data Research
- Infra at tremendous scale
or just generally want to know more about what we do @ProximalHQ, please reach out, I would love to chat :)
- Post-Training
- Data Research
- Infra at tremendous scale
or just generally want to know more about what we do @ProximalHQ, please reach out, I would love to chat :)
Excited to share i’ve joined @ProximalHQ here in SF! I think that there is a ton of interesting open problems related to data and post training at large. There is no better team to work with than the one we have and I am extremely excited to share our work :)
15
14
470
22,444
Many cybersecurity evals use differential execution to test whether agents can reproduce known software vulnerabilities
Without added constraints, this creates fairness issues: Our evaluation shows that CyberGym is effectively saturated on a verified task subset
Without added constraints, this creates fairness issues: Our evaluation shows that CyberGym is effectively saturated on a verified task subset
10
12
77
21,067
For more reliable cybersecurity evals, most robust graders are required:
SEC-Bench Pro achieves high precision/recall on vuln-repro tasks with LLM-as-judge, and ExploitGym additionally uses a Capture-the-Flag approach for clear certainty
We consider these benchmarks high sing them for cybersecurity evaluations
SEC-Bench Pro achieves high precision/recall on vuln-repro tasks with LLM-as-judge, and ExploitGym additionally uses a Capture-the-Flag approach for clear certainty
We consider these benchmarks high sing them for cybersecurity evaluations
1
0
7
461