Agent Economies and Gemma 4 QAT #42

June 9, 2026 — 2 min read

Today's Letter

Hugging Face published a June 8 field note from its Build Small Hackathon on a simulated economy run by five different small models
The rebuild replaced one model playing five roles with a mixed council: one OpenAI model, one NVIDIA model, one OpenBMB model, and a self-fine-tuned half-billion-parameter model running two agents
In the earlier single-model setup, a bank-run scenario pushed the honey price from 10 to 3, but the same setup did not reproduce after the model population changed
In the five-model run, agents reacted to the panic signal by hoarding honey instead of selling it, so the price rose on scarcity and the short trade lost money
The author reports three failed attempts to restore the crash through rumor, inventory oversupply, and a larger short position, with losses of 15, 26, and 27 pebbles
The post argues that shocks to supply or narrative only bias agent decisions, and that heterogeneous model populations can reject the behavior seen in a simpler test setup
The eventual fix was a settlement-stage override that directly halved the reference price after market clearing, making the bank-run outcome deterministic and turning the trade profitable
The main takeaway is a split between emergence and authored control: emergent behavior adds texture, while outcomes that must happen need an explicit control seam downstream of agent decisions

Source: huggingface.co

Google released Gemma 4 Quantization-Aware Training checkpoints across multiple model sizes for local deployment.
The update focuses on lower memory use while preserving model quality during quantized inference.
Google also introduced a mobile-oriented quantization format aimed at smaller on-device runtimes.
E2B said the new checkpoints can run in about ~1GB memory in some deployment setups.
Ecosystem support arrived immediately through Ollama and vLLM, reducing integration lag for local serving.
A conversion caveat also surfaced: naive export from QAT to llama.cpp Q4_0 can reduce accuracy.
Unsloth said its dynamic GGUF path recovers much of that lost accuracy during conversion.
The release stands out as an AI-model update with direct impact on local inference efficiency and portability.

Source: androidauthority.com
More: patmcguinness.substack.com

Meta said it disrupted an NSO-linked spearphishing campaign that targeted WhatsApp users and asked a US federal court to hold NSO Group in contempt.
The dispute follows Meta's 2019 lawsuit over Pegasus-related attacks on activists, journalists, political dissidents, and other users.
A jury awarded Meta $167 million in damages in the case, later reduced by a judge to $4 million, alongside a permanent injunction barring NSO from targeting WhatsApp or its users.
Meta said the new campaign used NSO-linked accounts and phishing domains designed to trick users into clicking malicious links tied to earlier NSO tactics.
The company said the latest campaign targeted fewer than 10 WhatsApp users, primarily in Jordan and Lebanon.
Meta said it found no signs of compromise among the identified targets and published the related domains so others can check for exposure on WhatsApp or other platforms.

Source: engadget.com
More: about.fb.com · techbuzz.ai · jpost.com

Jocoletter curates AI, software, and product trends for developers and builders.

#Google #HuggingFace #Meta