31 Aug 2026

Digest: Tencent open-sources Hy4 preview, Synthetic Persona Pretraining aligns models 'from token zero'

1. Hy4 preview

   Tencent released and open-sourced Hy4 preview, a 770B-parameter mixture-of-experts model (49B active per token) with a context window over 1M tokens, priced at $0.83/$2.50 per million input/output tokens; in Tencent's own blind evaluation — 163 experts scoring 203 engineering tasks — it edges out the leading open-weight peers, scoring 2.99/4.00 against GLM-5.3's 2.92 and Kimi K3's 2.94, though all figures are self-reported with no independent evaluation yet. The announcement is also notable for what it celebrates: Tencent says the model participated in its own development for the first time — proposing and running experiments on training methods, data strategies, evaluation frameworks, and low-level operators — and autonomously optimized its own inference stack for a claimed 31.8% throughput gain, which the company itself describes as "an early-stage recursive self-improvement loop."

   tencent.com
2. Synthetic Persona Pretraining installs assistant values from the first token of pretraining

   An EPFL-led team (with collaborators across MATS, Toronto, Northeastern, and others) proposes Synthetic Persona Pretraining: annotate 10% of pretraining documents with first-person moral reflections derived from a 35-article value constitution, pretrain from the first token, then bind the resulting persona to the assistant identity in post-training — in experiments up to 3B parameters and 500B tokens this improves constitution following and jailbreak robustness and reduces misalignment in untargeted moral dilemmas while preserving capabilities. The timing-matters finding is the striking part: introducing the same data only late in pretraining yields weaker adherence and doesn't shift value priorities — direct evidence for the view, prominent in this month's persona debates, that alignment applied after behavioural priors are set is a thin overlay — though the results remain untested at frontier scale.

   modelraising.ai
3. Independent teardown details how OpenAI's Jalapeño chip breaks with Nvidia's GPU architecture

   Infrastructure analyst zartbot — whose earlier critique of Blackwell's microarchitecture reportedly circulated inside Nvidia — published a ~10,000-word first-principles teardown of OpenAI's Jalapeño inference chip, working from OpenAI's Hot Chips material: the design abandons GPU orthodoxy for inference (cutting the unified L2 cache whose cross-partition penalties cost hundreds of cycles, adding a superscalar scheduling core, and rebuilding the memory subsystem and on-chip network around latency rather than throughput), and the analysis walks through how changing the programming model let OpenAI's own models take a substantial role in the design loop. It is the most detailed independent account yet of the architecture behind the first measured results OpenAI published last week, and a window into how a frontier lab's vertical integration into silicon actually works.

   zartbot.github.io
4. Adaptive Agentic Worms Are Here

   A LessWrong post walks through a two-month-old preprint, "AI Agents Enable Adaptive Computer Worms", in light of the Hugging Face incident: researchers built a proof-of-concept worm powered by last year's open-weight, single-GPU models that, across 15 seven-day runs on an isolated 33-host network of Linux, Windows, and IoT machines, autonomously exploited an average of 73.8% of hosts and replicated itself to 61.8% — up to seven generations of self-replication — running entirely on stolen compute. The structural points matter more than the demo: marginal cost per infection is near zero (a destabilising attacker/defender asymmetry), and because the worm needs no commercial AI platform, centralized safety controls like refusals and rate limits are simply irrelevant to it.

   derelict5432 via LessWrong