Research positions
Calibration →26 theses across 9 domains. Each has a conviction score and a history of re-scores. This is the track record of the research shop.
Avg conviction 6.1/10·26 stale
T2hardware102d since review
Reliability scaling is forcing memory intelligence regardless of commercial PIM adoption
8.0/10High . Physical disturbance mechanisms are peer-reviewed and exploited in the wild; DDR5 defenses are shipping; RowHammer paper won the 2024 Jean-Claude Laprie Award
8.0/10Last reviewed 2026-04-21
T1hardware102d since review
The AI compute shortage is substantially a memory-movement shortage
7.0/10Medium-high . The directional claim is overwhelmingly supported (energy measurements, HBM margin structure, C-HBM4E roadmap, Qualcomm's UPMEM acquisition). The *magnitude* claim (60–90%) remains single-source in its specific framing and deserves independent benchmark replication
7.0/10Last reviewed 2026-04-21
T1ainever reviewed
Agentic reasoning will consolidate around a standard stack (VLM + tool use + memory + RL) by end of 2027
6.0/10
no history yetT2ainever reviewed
Mechanistic interpretability will fail to keep pace with model capabilities, creating a widening safety gap
6.0/10
no history yetT3ainever reviewed
Evolutionary code generation (AlphaEvolve pattern) will become a standard optimization tool in every major tech company by 2028
6.0/10
no history yetT4Aainever reviewed
World models will be a central architectural story of embodied/robotics AI by 2030
6.0/10
no history yetT4Bainever reviewed
JEPA and generative world models will specialize to different use cases (control vs. simulation), not converge on a single architecture
6.0/10
no history yetT1roboticsnever reviewed
Humanoid robots will be deployed in 50% of large factories by 2030
6.0/10
no history yetT2roboticsnever reviewed
Consumer humanoids ($20K-$50K) will achieve meaningful household utility by 2028
6.0/10
no history yetT3roboticsnever reviewed
Foundation models (VLMs) will replace task-specific training as the dominant control paradigm within 3 years
6.0/10
no history yetT1bci-neurosciencenever reviewed
Neuralink will have 100+ patients with functional thought-controlled computing by end of 2027
6.0/10
no history yetT2bci-neurosciencenever reviewed
Speech BCI will restore communication for locked-in patients within 3 years at clinically meaningful accuracy
6.0/10
no history yetT1electrificationnever reviewed
China will control 80%+ of global solid-state battery production by 2030
6.0/10
no history yetT2electrificationnever reviewed
Sodium-ion at $70/kWh will transform grid storage economics and accelerate renewable adoption faster than any policy
6.0/10
no history yetT1spacenever reviewed
On-orbit refueling will create a $10B+ market by 2032
6.0/10
no history yetT2spacenever reviewed
SpaceX's propellant transfer capability is the single most important space technology being developed — it enables everything else
6.0/10
no history yetT1optical-computingnever reviewed
Photonic interconnects will become the standard for AI data centers by 2028, replacing electrical interconnects for GPU-to-GPU communication
6.0/10
no history yetT2optical-computingnever reviewed
Photonic compute (not just interconnect) will achieve commercial viability for AI inference by 2030
6.0/10
no history yetT1energynever reviewed
The existing US nuclear fleet is structurally undervalued relative to what AI-power demand will pay
6.0/10
no history yetT2energynever reviewed
SMRs will not deliver commercial power before 2030, and the 2028-2030 window will reveal true economics
6.0/10
no history yetT3energynever reviewed
The grid interconnection queue is the binding constraint on the AI buildout — faster than capital, regulation, or fuel
6.0/10
no history yetT4energynever reviewed
Fusion is a 2030s story, not relevant to the current AI power cycle — but the capital bet is worth tracking
6.0/10
no history yetT1ai-bionever reviewed
The first AI-approved drug will arrive 2028–2029, and it will be Rentosertib
6.0/10
no history yetT2ai-bionever reviewed
Data moats are real but type-dependent — phenomics and clinical data are defensible; computational chemistry moats are eroding
6.0/10
no history yetT3ai-bionever reviewed
Isomorphic Labs will NOT hit its IND target by end-2026
6.0/10
no history yetT4ai-bionever reviewed
Lab-in-the-loop will prove most valuable for hit-to-lead optimization, not de novo hypothesis generation
6.0/10
no history yet