Rising compute barriers: Researchers complain reviewers demand evals on larger models
chhaviyadav_ · x · 2026-07-26
A researcher reminisced about getting into ICLR with 12-hour experiments on a GTX 1060, complaining that arXiv and top conferences are increasingly gatekeeping newcomers.
Another highlighted that academia has moved past 8B models, noting that reviewers now explicitly demand evaluations and fine-tuning on larger, advanced open-source models like the 21B GPT-OSS. This sparked a discussion on which open models are currently feasible for academic papers.
More from AGI Musings
- A Reddit thread argues open-weight models mainly democratize defense, not danger — enginetown · 2026-07-26
- Andreessen says AI coding can make engineers 20x faster — and could still trigger layoffs — r0ck3t23 · 2026-07-26
- If AI models really commoditize, Google may be right to hold back its SOTA model — cocktailpeanut · 2026-07-26
- Conversation links human reasoning, AI agents and research infrastructure for more reliable science — himanshustwts · 2026-07-26
- AI agents may need a knowledge-graph “highway system” on top of the web — CyborgWriter · 2026-07-26
- Bindu Reddy says open source is the safest path for AI, not bans — bindureddy · 2026-07-26