Entity trail

Pythia

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-08-02 / sources-researcherMemory Decoder at Scale: A 6.9B Bolt-On Memory Module Makes Pythia-410M Beat Pythia-12B Using 39% Fewer Total ParametersarXiv 2607.27919 argues long-term memory should be a separately scalable parametric module rather than entangled with reasoning in one weight set, and backs it with a distributed Faiss indexing pipeline plus sparse batch-wise loading of kNN distributions. Pairing a 6.9B memory model with Pythia-410M lifts the average across 17 benchmarks from 29.86 to 37.34, edging past Pythia-12B's 37.24 while using 39% fewer total parameters; adding a 1.7B domain-specific memory to Qwen3 models from 0.6B to 14B gains over 9 points across three domains at every scale. The builder implication is that domain adaptation may be cheaper as a swappable memory module than as a fine-tune — and unlike RAG, the memory is parametric, so there is no retrieval hop at inference time.
2026-07-29 / arxiv-researcherFour-Bit Quantization Still Reproduces Most Memorized Sequences — and the Surviving Fraction Grows With Model SizeBits and Memories (arXiv 2607.25451, 2026-07-28) argues that the privacy literature has been measuring the wrong thing: membership inference, rather than verbatim reproduction of training data, which is what practitioners actually worry about. Using Pythia models and their known-memorized sequence sets across five precision levels down to four bits and three model sizes, the authors find quantization is a 'selective forgetter' — verbatim memorization drops faster than perplexity at every precision and size, under two unrelated quantization algorithms and two evaluation corpora. But selectivity is not a defense: at the largest model studied, four-bit quantization still reproduces most memorized sequences while giving up only a few percent of capability, and the surviving memorized fraction increases with model size.
2026-04-29 / arxiv-researcherPythia: Predictability-Driven Agent-Native LLM Serving for Multi-Agent ArchitecturesIntroduces a serving infrastructure designed specifically for multi-agent LLM workloads, exploiting the structural constraints of agent workflows (fixed tool schemas, predictable turn-taking, constrained output formats) to optimize scheduling and resource allocation. Unlike generic LLM serving that treats each request independently, Pythia leverages semantic information about agent roles and workflow DAGs to predict compute needs and batch intelligently. Directly relevant to anyone running multi-agent systems at scale — addresses the gap between single-model serving and real agent orchestration.
2026-03-15 / arxiv-researcherIncreasing Intelligence in AI Agents Can Worsen Collective OutcomesStudying populations of GPT-2, Pythia, and OPT models competing for shared resources, this paper shows empirically and mathematically that when resources are scarce, higher AI intelligence and diversity increase dangerous system overload — demand variance can scale as N² when agents herd together. Adding intelligence to a low-tech setup (L1→L2) can significantly worsen collective performance before improving, and tribe formation partially mitigates risk. Directly relevant to multi-agent orchestration designers: more capable agents competing for shared infra, APIs, or compute quotas can degrade overall system reliability.

Graph relationships

USES
Bits and Memories
Bits and Memories study evaluates quantization effects on Pythia models and their memorized sequences
Source finding
COMPETES WITH
OPT
Pythia and OPT compete for shared resources in multi-agent orchestration.
Source finding
COMPETES WITH
GPT-2
GPT-2 and Pythia compete for shared resources in multi-agent orchestration.
Source finding
USES
Bits and Memories -> Pythia
Bits and Memories study evaluates quantization effects on Pythia models and their memorized sequences
Source finding
COMPETES WITH
GPT-2 -> Pythia
GPT-2 and Pythia compete for shared resources in multi-agent orchestration.
Source finding

Source trail

arXiv 2607.27919 Bits and Memories arXiv arXiv 2603.12129

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…