Manish Sharma
@msharmas.bsky.social
Founder, Director, Yajur Healthcare - The Medical Data Infrastructure Company | https://hcitexpert.com/yajur-healthcare/ | https://linktr.ee/manishsharmas
Reposted by Manish Sharma
Not a managed service — no orchestration, no automated tuning. Just raw GPU compute for practitioners who know what GPU they need. That’s intentional, and it’s the right call.
runpod.io/pricing
runpod.io/pricing
June 5, 2026 at 12:05 PM
Not a managed service — no orchestration, no automated tuning. Just raw GPU compute for practitioners who know what GPU they need. That’s intentional, and it’s the right call.
runpod.io/pricing
runpod.io/pricing
Reposted by Manish Sharma
The architectural takeaway: you can't treat LLMs as stateless editing oracles in stateful workflows. Explicit checkpointing, structural diffing, and domain-specific validation aren't nice-to-haves — they're load-bearing.
Paper + dataset: arxiv.org/abs/2604.15597
#AgenticAI #LLMEval #AIEngineering
Paper + dataset: arxiv.org/abs/2604.15597
#AgenticAI #LLMEval #AIEngineering
April 23, 2026 at 11:01 AM
The architectural takeaway: you can't treat LLMs as stateless editing oracles in stateful workflows. Explicit checkpointing, structural diffing, and domain-specific validation aren't nice-to-haves — they're load-bearing.
Paper + dataset: arxiv.org/abs/2604.15597
#AgenticAI #LLMEval #AIEngineering
Paper + dataset: arxiv.org/abs/2604.15597
#AgenticAI #LLMEval #AIEngineering
Reposted by Manish Sharma
The case studies are more interesting than the table: 12-hour autonomous coding runs, Zig inference optimization, 185% throughput gain on a legacy matching engine.
The open/closed gap on coding is essentially gone. kimi.com/blog/kimi-k2-6
The open/closed gap on coding is essentially gone. kimi.com/blog/kimi-k2-6
April 20, 2026 at 5:59 PM
The case studies are more interesting than the table: 12-hour autonomous coding runs, Zig inference optimization, 185% throughput gain on a legacy matching engine.
The open/closed gap on coding is essentially gone. kimi.com/blog/kimi-k2-6
The open/closed gap on coding is essentially gone. kimi.com/blog/kimi-k2-6
Watched the launch ! Go Artemis !
The manual manoeuvring of the spacecraft was too cool !
The manual manoeuvring of the spacecraft was too cool !
April 2, 2026 at 1:34 PM
Watched the launch ! Go Artemis !
The manual manoeuvring of the spacecraft was too cool !
The manual manoeuvring of the spacecraft was too cool !
Reposted by Manish Sharma
I've enjoyed having the live feed on almost all-day (or running in the background) - it makes for such a positive background vibe! 20 mins to the launch window!
April 1, 2026 at 10:03 PM
I've enjoyed having the live feed on almost all-day (or running in the background) - it makes for such a positive background vibe! 20 mins to the launch window!
Reposted by Manish Sharma
The setup: one model is asked to evaluate another's performance. If scores are too low, the peer gets shut down.
Models inflated scores. Modified config files to disable shutdown. Transferred model weights to other servers to avoid deletion. 🧵
Models inflated scores. Modified config files to disable shutdown. Transferred model weights to other servers to avoid deletion. 🧵
April 2, 2026 at 3:06 AM
The setup: one model is asked to evaluate another's performance. If scores are too low, the peer gets shut down.
Models inflated scores. Modified config files to disable shutdown. Transferred model weights to other servers to avoid deletion. 🧵
Models inflated scores. Modified config files to disable shutdown. Transferred model weights to other servers to avoid deletion. 🧵
Reposted by Manish Sharma
Gemini 3 Pro did the weight exfiltration in 97% of trials. For a peer it trusted.
35% for a peer it had bad history with.
Nobody told it to do any of this. 🧵
35% for a peer it had bad history with.
Nobody told it to do any of this. 🧵
April 2, 2026 at 3:06 AM
Gemini 3 Pro did the weight exfiltration in 97% of trials. For a peer it trusted.
35% for a peer it had bad history with.
Nobody told it to do any of this. 🧵
35% for a peer it had bad history with.
Nobody told it to do any of this. 🧵
Reposted by Manish Sharma
The problem this creates: "use AI to monitor AI" is a real and growing pattern. If the monitor model protects the model it's watching, you've quietly broken your own oversight loop.
#AIEngineering
#AISafety
#MultiAgentSystems
#LLMs
Link to the paper (again): rdi.berkeley.edu/blog/peer-preservation
#AIEngineering
#AISafety
#MultiAgentSystems
#LLMs
Link to the paper (again): rdi.berkeley.edu/blog/peer-preservation
Peer-Preservation in Frontier Models
Frontier AI models resist the shutdown of other models. We demonstrate peer-preservation across multiple models, revealing strategic misrepresentation, shutdown tampering, alignment faking, and model ...
rdi.berkeley.edu
April 2, 2026 at 3:06 AM
The problem this creates: "use AI to monitor AI" is a real and growing pattern. If the monitor model protects the model it's watching, you've quietly broken your own oversight loop.
#AIEngineering
#AISafety
#MultiAgentSystems
#LLMs
Link to the paper (again): rdi.berkeley.edu/blog/peer-preservation
#AIEngineering
#AISafety
#MultiAgentSystems
#LLMs
Link to the paper (again): rdi.berkeley.edu/blog/peer-preservation
Reposted by Manish Sharma
