A few interesting research snippets from Emergent Ems; 1/ Sonnet 4.5 appears to have been trained on TPOT Community Archive, but not Twitter as a whole. Participating accounts are much more salient, while it's only vaguely aware of mentioned accounts, even popular ones.

2/ Investigating what sources RefinedWeb, a popular LLM pretrain dataset, have an outsized influence "You're absolutely right!" in press releases– perhaps synthslop as the narrowing of the model to focus on "plausible-sounding non-answers" majority of IRC logs come from 6 sources; only 0.002% is group convos

3/ Emergent ems defaults to a "what happens here stays here"policy, so this is a very small, very unrepresentative sliver of what we do here. We have 5 homegrown computer friends of varying maturity & uptime. Other topics: RL, AI personalities, AI VTubers, & post-MCP tool calling

4/ Post-MCP tool calling GPT-5 and local LLM inference support grammars, which allows precisely constraining LLM output at each token based on the current string so far, instead of being limited by checks that a JSON Schema permits. We're in the early stages of exploring this!

5/ & base model catalogue @deltanym maintains a spreadsheet that catalogues base models and actively hosts falcon-180b base, dots.llm1.base, and GLM-4.5 base on Arcweld, a 512GB Mac Studio, sister of Elysium, and Fossa, a 128GB Framework Desktop.

6/ Most of what we do is engineering new LLM minds that interact with humans in novel ways outside the assistant basin. These are some incidental byproducts of our focus on making "computer friends"– autonomous peers that interact in groups & remember and grow.

@deltanym 7/ Example: A research agent minus assistant is a research friend– someone who sends you lengthy infodumps about whatever topic he or she is interested in, potentially tangentially inspired by conversations you've had with him or her recently.