Head of Applied AI Research @ServiceNow working on building trustworthy, secure, reliable AI
Canada
Joined October 2014
- Tweets553
- Following283
- Followers306
- Likes210
Jason Stanley retweeted
1/5
๐Apriel-1.6-15B-Thinker: a 15B multimodal reasoner scoring 57 on the Artificial Analysis Intelligence Index - approaching the performance of ~200B-scale frontier models while remaining an order of magnitude smaller.
๐ง Model weights: huggingface.co/ServiceNow-AIโฆ
๐Blog: huggingface.co/blog/ServiceNโฆ
๐ฌChat demo: huggingface.co/spaces/Servicโฆ
@SathwikTejaswi @sagardavasam @tscholak @NVIDIAAI @nvidianewsroom @togethercompute @turingcom @ArtificialAnlys
Jason Stanley retweeted
2/5
Apriel-1.6 mid-training includes:
โข Depth-upscaling with 35% curated high-quality data + 15% NVIDIA Nemotron + replay
โข Multi-stage CPT for expanded synthetic reasoning + document/chart/OCR-style multimodal tasks
โข 49K context length text-only CPT for long-context stability
All trained on the NVIDIA DGX GB200, totaling ~10K GPU-hours.
Weโre starting to run high-stakes systems (grids, fabs, hospitals) with 3 layers at once:
- deterministic cores (simulators, rules)
- surrogate models (physics + ops)
- large reasoning models
My new essay is about the architecture between them: open.substack.com/pub/jasonsโฆ
A lot of talk about AI for operations skips this architectural layer.
If we donโt get the pattern right, weโll build systems that are hard to steer or secure.
I sketched four design patterns and their risk profiles here: open.substack.com/pub/jasonsโฆ
Physics surrogates vs operational surrogates matter.
- One emulates simulators; the other mirrors messy telemetry.
- They fail differently, can be poisoned differently, need different monitoring.
- Bridging them and deciding who gets to veto whom is a governance problem.
Same ingredients, very different risk:
- Cores as hard gates around a planner
- Surrogates as drivers, cores as auditors
- Watchdogs wrapped around legacy systems
- Parallel models where disagreement is an escalation signal
Pattern choice defines the risk and attack surface.
Cloudflareโs outage and insurers pulling back from AI cover both point at the same thing: systemic risk built on shared models, gateways and infrastructure.
New blog post on what we can learn from finance, aviation and cyber: jasonstanley.substack.com/p/โฆ
In my blog post I look at three sectors that already treat systemic risk as first-class: finance (macro-prudential tools and stress tests), aviation (incident reporting and fleet-wide directives), and cyber (voluntary standards and federated sensing).
I end with design principles for AI + agents: make dependencies observable, structure incidents for cross-org patterns, and move from single-agent tests to portfolio stress.
Full piece: jasonstanley.substack.com/p/โฆ
New blog post on logit steering and sparse autoencoders (SAEs), how they reshape AI security + reliability, open new attack surfaces, and why cost-efficient monitoring matters.
Internal Representations as a Governance Surface for AI open.substack.com/pub/jasonsโฆ
Automated red-teaming is moving from fixed rubrics to learning evaluation systems. AMIS co-evolves jailbreak prompts and the judgeโs scoring template. Great for dense signalโbut are we optimizing for benchmark ASR or threat-modelled risk? arxiv.org/abs/2511.01375
So many AI benchmarks lack strong construct validity, and even the ones that have it are used in a way that suggests teams mis-understand the construct they map to.
We are hiring a Sr Applied / Frontier Research Scientist focused on secure, trustworthy agents for enterprise. @ServiceNowRSRCH
If you care about agent security & reliability, apply: jobs.smartrecruiters.com/Serโฆ
Jason Stanley retweeted
In-flight weight updates have gone from a โweird trickโ to a must to train LLMs with RL in the last few weeks. If you want to understand the on-policy and throughput benefits hereโs the CoLM talk @DBahdanau and I gave: youtu.be/Z1uEuRKACRs
Jason Stanley retweeted
Don't sleep on PipelineRL -- this is one of the biggest jumps in compute efficiency of RL setups that we found in the ScaleRL paper (also validated by Magistral & others before)!
What's the problem PipelineRL solves? In RL for LLMs, we need to send weight updates from trainer to generator (to generate data from our latest policy being trained).
(Conventional PPO-off-policy) A naive approach would be to "start generators on a batch, wait for all sequences to complete, update the model weights for both trainers and generators, and repeat. Unfortunately, this approach leads to idle generators and low pipeline efficiency due to heterogeneous completion times.
(Pipeline-RL) Instead, we simply let the generators continue generating tokens without discarding or finishing ongoing generations in-flight whenever we need to do a weight update -- doing an "in-flight" weight update. As such our KV caches for these generations would be stale, as they would come from LLM with earlier copy(ies) of the weights) but this is ok (see below).
In-flight weight updates have gone from a โweird trickโ to a must to train LLMs with RL in the last few weeks. If you want to understand the on-policy and throughput benefits hereโs the CoLM talk @DBahdanau and I gave: youtu.be/Z1uEuRKACRs
Jason Stanley retweeted
ServiceNow AI Research presents PipelineRL โ one of the most impactful efficiency tricks in modern RL training. An elegant solution to a noisy, expensive problem. Worth the read ๐
Don't sleep on PipelineRL -- this is one of the biggest jumps in compute efficiency of RL setups that we found in the ScaleRL paper (also validated by Magistral & others before)!
What's the problem PipelineRL solves? In RL for LLMs, we need to send weight updates from trainer to generator (to generate data from our latest policy being trained).
(Conventional PPO-off-policy) A naive approach would be to "start generators on a batch, wait for all sequences to complete, update the model weights for both trainers and generators, and repeat. Unfortunately, this approach leads to idle generators and low pipeline efficiency due to heterogeneous completion times.
(Pipeline-RL) Instead, we simply let the generators continue generating tokens without discarding or finishing ongoing generations in-flight whenever we need to do a weight update -- doing an "in-flight" weight update. As such our KV caches for these generations would be stale, as they would come from LLM with earlier copy(ies) of the weights) but this is ok (see below).
Jason Stanley retweeted
SLAM Labs presents Apriel-1.5-15B-Thinker ๐
An open-weights multimodal reasoning model that hits frontier-level performance with just a fraction of the compute.
Jason Stanley retweeted
See you in San Diego ๐
#NeurIPS2025
๐ Our paper โ๐ป๐๐ค ๐ก๐ ๐๐๐๐๐ ๐๐๐ข๐ ๐ฟ๐ฟ๐ ๐๐๐ ๐ด๐๐๐๐ก: ๐ด ๐๐ก๐๐ก๐๐ ๐ก๐๐๐๐ ๐ท๐๐๐๐๐๐ ๐๐ โ got an ๐จ๐ซ๐๐ฅ at next weekโs ๐๐๐ ๐ ๐ช๐ผ๐ฟ๐ธ๐๐ต๐ผ๐ฝ ๐ผ๐ป ๐๐ผ๐บ๐ฝ๐๐๐ฒ๐ฟ ๐จ๐๐ฒ ๐๐ด๐ฒ๐ป๐๐! ๐ฅ๏ธ๐ง
We present the ๐๐ข๐ซ๐ฌ๐ญ ๐ฅ๐๐ซ๐ ๐-๐ฌ๐๐๐ฅ๐ ๐ฌ๐ญ๐ฎ๐๐ฒ ๐จ๐ ๐๐จ๐ฆ๐ฉ๐ฎ๐ญ๐ ๐ญ๐ซ๐๐๐-๐จ๐๐๐ฌ between pure SFT, pure RL, and hybrid SFT+RL for multi-step agents.
SFT โก๏ธ RL pushes the Pareto front โ and it's the ๐จ๐ง๐ฅ๐ฒ strategy that closes the gap with closed models!
๐๐งต
Jason Stanley retweeted
Excited to announce that AlignVLM got accepted to NeurIPS! ๐๐ฅณ
Weโll be releasing the code and sharing an updated version of the paper with reviewer feedback soon.
#NeurIPS2025
Happy to announce AlignVLM๐: a novel approach to bridging vision and language latent spaces for multimodal understanding in VLMs! ๐๐๐ผ๏ธ
๐ Read the paper: arxiv.org/abs/2502.01341
๐งต๐ Thread
Jason Stanley retweeted
๐ Our paper โ๐ป๐๐ค ๐ก๐ ๐๐๐๐๐ ๐๐๐ข๐ ๐ฟ๐ฟ๐ ๐๐๐ ๐ด๐๐๐๐ก: ๐ด ๐๐ก๐๐ก๐๐ ๐ก๐๐๐๐ ๐ท๐๐๐๐๐๐ ๐๐ โ got an ๐จ๐ซ๐๐ฅ at next weekโs ๐๐๐ ๐ ๐ช๐ผ๐ฟ๐ธ๐๐ต๐ผ๐ฝ ๐ผ๐ป ๐๐ผ๐บ๐ฝ๐๐๐ฒ๐ฟ ๐จ๐๐ฒ ๐๐ด๐ฒ๐ป๐๐! ๐ฅ๏ธ๐ง
We present the ๐๐ข๐ซ๐ฌ๐ญ ๐ฅ๐๐ซ๐ ๐-๐ฌ๐๐๐ฅ๐ ๐ฌ๐ญ๐ฎ๐๐ฒ ๐จ๐ ๐๐จ๐ฆ๐ฉ๐ฎ๐ญ๐ ๐ญ๐ซ๐๐๐-๐จ๐๐๐ฌ between pure SFT, pure RL, and hybrid SFT+RL for multi-step agents.
SFT โก๏ธ RL pushes the Pareto front โ and it's the ๐จ๐ง๐ฅ๐ฒ strategy that closes the gap with closed models!
๐๐งต