
AI @ServiceNow | ai rl code agents | endurance athlete | co-founded @aikynetix
Silicon Valley, CA
Joined August 2016
- Tweets3.8K
- Following1.6K
- Followers817
- Likes11.5K
Pinned Tweet
Excited to present at the Berkeley Agentic AI Summit this week: “Are Small LLMs Ready for Coding Agents?”
rdi.berkeley.edu/events/agen…
@UCBerkeley @berkeley_ai
#ai #agents #codingagents #berkeley #LLM
1
6
318
Does anyone else feel like Fable 5.2 was really disappointing? Back to Opus 55 Max and GPT6 Astra Ultra 🤷♂️
2
5
171
Today I discovered my first bug in LLVM compiler and IR code generation. This bug is now confirmed:
github.com/llvm/llvm-project…
1
2
127
Denis A. retweeted
🚀 Meet Apriel Voice 27B Preview: an open-weight voice AI built to get work done.
Beyond conversation: adapt to dynamic instructions, reason across modalities, and orchestrate multi-step tool calls to complete real tasks.
See Apriel Voice in action. 👇
Pystachy runtime and compiler for Python now runs @karpathy NanoGPT and MicroGPT training loops 10x to 100x faster with very little changes:
github.com/den-run-ai/pystac…
5
104
Cuyamaca 💯 km - first 100k ultramarathon, WS100 qualifying time
@wser
1
4
140
Python is a simple and expressive language. But runtime overhead remains a major barrier for many CPU-intensive workloads.
Pystachy compiles a statically typed subset of Python to fast native code while retaining Python readability and reusing existing CPython library code.
Compatibility checks include 700+ Python programs in Pystachy compared with CPython through JIT and AOT. Several standard-library ports need only a few localized changes - the adapted code remains valid Python and can be diffed against upstream.
15
Denis A. retweeted
Agentic post-training needs data that targets the model's failure modes, so you teach it what it doesn't know yet. One failure tells you something, but training needs many feasible, verifiable tasks that exercise the same capability in different situations.
We built AutoSynthData to turn those capability gaps into training data. It uses a target model's failures and a stronger teacher's successes to decide what the model should learn next, then generates and validates new tasks that exercise those capabilities. As the model improves, the curriculum shifts toward what it still finds difficult.
On EnterpriseOps Gym (Hybrid), 2,000 generated tasks raised Gemma 4 26B's Pass@1 by 7.2 points (35% relative).
Read our full blog here: huggingface.co/blog/ServiceN…
w/ Shruthan, @den_run_ai, @sagardavasam
This weekend I am cycling ADA Tour De Cure 100k with Team @ServiceNow
This ride is dedicated to all people fighting against diabetes, especially my mom, and some of my close relatives.
Please donate at: tour.diabetes.org/participan…
@AmDiabetesAssn
#cycling #fundraising #diabetes
1
6
398
Denis A. retweeted
Paper: paperswithcode.co/paper/2608…
A smaller open-weight model with an evolved harness beat GPT-5.5 on the default harness by 19.2 points on ITBench. The right scaffold can matter more than model size.
Denis A. retweeted
StarHarness: Evolve the harness, not the model
ServiceNow's new method searches for the optimal agent scaffold around a fixed model, boosting performance by 20–35 points across enterprise benchmarks while cutting inference cost up to 53%.
Denis A. retweeted
On ITBench, Qwen3.5-27B with an evolved harness beats GPT-5.5 on the baseline harness by +19.2 points.
More capable models do not always make better enterprise agents.
With the right harness, a smaller open-weight model can outperform a larger frontier model running the default scaffold.
In our new preprint, we present StarHarness, a method for directly searching for an environment-optimal harness around a fixed model.
Across three enterprise benchmarks, the evolved harness improved performance by 20–35 percentage points and reduced inference cost by 17–53%.
Our results suggest that some of the capability gap we attribute to models may actually come from how well the surrounding harness fits the environment.
Read the full preprint here: arxiv.org/abs/2608.24804
w/@den_run_ai, Vikas Yadav, @RajeswarSai, @patricebechard, @siris_krishna, @sagardavasam
New vibecoding competition from @Meta: build a Three.js mobile game prototype in 3 weeks. $300k in prizes.
mhcp-game-prototype.devpost.…
















