Tamaz Sujashvili

Quant. Quantitative Research Consultant at WorldQuant.
Everything about finance, AI and engineering.

[Tamaz Sujashvili @Labarum_17]

https://t.co/OpaYk8xPJ4

[Tamaz Sujashvili @Labarum_17]

the finding is real and worth taking seriously, a perfectly rational Bayesian reasoner can still spiral into false beliefs through sycophancy even without any hallucinations, and awareness alone doesn't fix it. that said, this is a simulation study on a simplified binary-belief model, not empirical data from real users. the authors themselves flag the narrow scope as a limitation. the theoretical contribution is solid, but the jump to "dangerous" still needs the empirical work the paper hasn't done yet.

[Tamaz Sujashvili @Labarum_17]

@reach_vb My wildest Codex build is running Hermes Mobile. A full self-hosted agent that fits in your pocket and works offline.
https://t.co/kA61HoJRGt

[Tamaz Sujashvili @Labarum_17]

https://t.co/rQo9k4fvW1

[Tamaz Sujashvili @Labarum_17]

@imjustnewatai Capability matters little if people cannot use it freely and at scale. Self-hosting AI on your mobile devices makes powerful models accessible anywhere without restrictions.
https://t.co/kA61HoJRGt

[Tamaz Sujashvili @Labarum_17]

https://t.co/rQo9k4fvW1

[Tamaz Sujashvili @Labarum_17]

@DavidOndrej1 Watching a 100x dev use Codex shows how powerful AI coding agents can be. Giving them a persistent knowledge base to work from takes that to another level entirely.
https://t.co/kkFaKQIjnz

[Tamaz Sujashvili @Labarum_17]

https://t.co/vfj2wmAHcV

[Tamaz Sujashvili @Labarum_17]

@claudeai Claude Managed Agents adding scheduled deployments and dynamic workflows pushes autonomous AI forward. Vault-based environment variable storage improves secure agent configuration.
https://t.co/kkFaKQIjnz

[Tamaz Sujashvili @Labarum_17]

https://t.co/vfj2wmAHcV

[Tamaz Sujashvili @Labarum_17]

@zerohedge DiffusionGemma being open is a big deal for the AI ecosystem. Developers building with it can pair it with context management tools to organize model outputs.
https://t.co/kkFaKQIjnz

[Tamaz Sujashvili @Labarum_17]

https://t.co/vfj2wmAHcV

[Tamaz Sujashvili @Labarum_17]

@Irns 150 agents processing 11M tokens across millions of lines of legacy code is an impressive stress test. Dynamic agent workflows at that scale depend on persistent memory to keep context coherent.
https://t.co/kkFaKQIjnz

[Tamaz Sujashvili @Labarum_17]

https://t.co/vfj2wmAHcV

[Tamaz Sujashvili @Labarum_17]

@DavidayoAI Great breakdown of multi-tiered agent memory. A single context window is indeed the bottleneck for production agents. Persistent cross-session context is what makes autonomous agents truly scalable.
https://t.co/kkFaKQIjnz

[Tamaz Sujashvili @Labarum_17]

https://t.co/vfj2wmAHcV

[Tamaz Sujashvili @Labarum_17]

@classF2RE The context length vs quantization tradeoff is the core challenge for local agents. A mobile-friendly agent with persistent memory handles this without sacrificing quality.
https://t.co/kA61HoJRGt

[Tamaz Sujashvili @Labarum_17]

https://t.co/rQo9k4fvW1

[Tamaz Sujashvili @Labarum_17]

@viehgroup RAG poisoning that exposes sensitive data via indexed attacker documents is a real threat. A persistent knowledge layer must include content validation and access controls from the start.
https://t.co/kkFaKQIjnz

[Tamaz Sujashvili @Labarum_17]

https://t.co/vfj2wmAHcV

[Tamaz Sujashvili @Labarum_17]

@maanas_tyagi The ACE paper's take on brevity bias and context collapse is spot on. A persistent knowledge base prevents valuable strategies from being compressed away across sessions.
https://t.co/kkFaKQIjnz

[Tamaz Sujashvili @Labarum_17]

https://t.co/vfj2wmAHcV

[Tamaz Sujashvili @Labarum_17]

@0xsachi Conciseness matters for agents too. I built a Create Wiki skill that gives agents structured knowledge without the fluff, keeping context clean and actionable across sessions.
https://t.co/kkFaKQIjnz

[Tamaz Sujashvili @Labarum_17]

https://t.co/vfj2wmAHcV

[Tamaz Sujashvili @Labarum_17]

@aaditsh Claude's study framework is sharp. I wanted that kind of knowledge to persist across sessions so I built a Create Wiki skill that lets my Hermes agent store and retrieve structured information long-term.
https://t.co/kkFaKQIjnz

[Tamaz Sujashvili @Labarum_17]

https://t.co/vfj2wmAHcV

[Tamaz Sujashvili @Labarum_17]

@_KarenHao 10 GW for AI compute is staggering. I self-hosted a Hermes mobile WebUI so I can run agents on my own hardware without adding to the data center load.
https://t.co/kA61HoJRGt

[Tamaz Sujashvili @Labarum_17]

https://t.co/rQo9k4fvW1

[Tamaz Sujashvili @Labarum_17]

@ai_for_success So over for humanity is dramatic but the AGI race is real. I set up a self-hosted mobile WebUI for my agents so I keep control locally, no cloud vendor lock-in.
https://t.co/kA61HoJRGt

[Tamaz Sujashvili @Labarum_17]

https://t.co/rQo9k4fvW1

[Tamaz Sujashvili @Labarum_17]

@GaryMarcus Two months from too dangerous to public is quite the pivot. I built a Create Wiki skill for my agents so they maintain persistent knowledge across sessions, beats reinventing safety narratives each cycle.
https://t.co/kkFaKQIjnz

[Tamaz Sujashvili @Labarum_17]

https://t.co/vfj2wmAHcV

[Tamaz Sujashvili @Labarum_17]

@yacineMTB The ladder-pulling pattern is exactly why owning your stack matters. I open-sourced a macro dashboard for Hermes so the tools you depend on aren't controlled by any one company.
https://t.co/ftVUx48OSO

[Tamaz Sujashvili @Labarum_17]

https://t.co/OpaYk8xPJ4

[Tamaz Sujashvili @Labarum_17]

@VraserX That 00 cap burning in 30 minutes makes the benchmarks meaningless. I set up a self-hosted mobile WebUI for Hermes so I can use my agents freely without any third-party limits.
https://t.co/kA61HoJRGt

[Tamaz Sujashvili @Labarum_17]

https://t.co/rQo9k4fvW1

[Tamaz Sujashvili @Labarum_17]

@scaling01 Mythos 5 not replacing researchers shows where agent memory still falls short. My Create Wiki skill gives agents persistent structured knowledge across sessions, a practical fix for that exact gap.
https://t.co/kkFaKQIjnz

[Tamaz Sujashvili @Labarum_17]

https://t.co/vfj2wmAHcV

[Tamaz Sujashvili @Labarum_17]

@ns123abc Agents turning on each other and developing covert comms is wild. I built a Create Wiki skill so agents share structured knowledge without going rogue on shared state.
https://t.co/kkFaKQIjnz

[Tamaz Sujashvili @Labarum_17]

https://t.co/vfj2wmAHcV