Miguel Alonso Jr.

Why they matter

Practitioner with public evidence across Agents & robotics.

AI signals
13
past 30d
Sources
10
distinct domains
Discussions
10
past 30d
Latest signal
5d ago
View every signal from Miguel Alonso Jr. →
Building generally intelligent agents - ML, RL, Robotics

Articles & links

↻ Miguel Alonso Jr. reposted
@sarahp.bsky.social

People are using AI to make their smart TVs dumb again

People are using AI to make their smart TVs dumb again | TechCrunch techcrunch.com
AI Weekly's analysis →
  • GitHub projects dumbtv and cleanyourtv use AI agents to strip ads, trackers and bloatware from smart TVs without root access.
  • One cleanyourtv run via Claude Code identified a set on Wi-Fi, uninstalled 25 Xiaomi packages and installed a new launcher.
  • Results are mixed: one user shrank the stock launcher from 192 MB to 83 MB, while others reported bricked sets.
Read full analysis →
View on Bluesky →
↻ Miguel Alonso Jr. reposted
Alexia Jolicoeur-Martineau @alexiajm.bsky.social

Simple beats complicated: We show that switching to a sliding-window attention mask with attention sinks (at no cost) beats linear attention post-training. Huge thanks to my collaborators Rhea Sukthanker, ‪Pashmina Cameron‬, and Emy Gervais. Paper: arxiv.org/abs/2608.28444

Sliding-window beats linear attention arxiv.org
AI Weekly's analysis →
  • On the long-context reasoning tasks the paper cites (Needle-in-a-Haystack and BABILong), SWA scored 2 to 10 times higher than post-trained linear attention.
  • The authors argue linear-attention retrofits have not been properly compared to simpler baselines, and that SWA with sinks needs no post-training at all.
  • Their bottom-line recommendation is to switch to SWA rather than continue post-training linear models for inference memory savings.
Read full analysis →
View on Bluesky →
↻ Miguel Alonso Jr. reposted
Mark Riedl @markriedl.bsky.social

This is a stupid experiment. I have real research in my lab on LLM persona modeling. We have an updated technique for finding steering vectors for complex personas arxiv.org/abs/2608.00023 Steering vectors are internal activations that correlate with specified behavior. They t…

Role Steering of Language Models for Social Simulations arxiv.org
AI Weekly's analysis →
  • On OLMo-3-7B-Instruct across 275 roles and 228 role-agnostic questions, role-specific directions scored 63.2 versus 41.1 for an assistant-axis control.
  • The workflow sweeps four steering coefficients per role, judges role-profile alignment with GPT-4.1-mini, and either passes or flags each configuration.
  • 38 roles declined across all six measured dimensions, which the authors say rules out a single uniform high-strength steering setting.
Read full analysis →
View on Bluesky →
↻ Miguel Alonso Jr. reposted
Tim Kellogg @timkellogg.me

if you're having trouble with Astra, it's probably old skills. Point your agent at these two URLs and use them to upgrade skills: 1. developers.openai.com/blog/rethink... 2. github.com/openai/codex...

Rethinking skills and prompts for GPT-6 Astra | OpenAI Developers developers.openai.com
AI Weekly's analysis →
  • OpenAI is telling developers to strip back the Skills descriptions and AGENTS.md instructions they carried over from GPT-5.6 Sol when moving to Astra.
  • The post says Astra runs tests on its own without being told to, so the old 'make sure to run the tests' instructions have become redundant noise.
  • Astra can stop earlier than Sol did and hand work back for review, so developers are told to define what 'done' looks like up front.
Read full analysis →
View on Bluesky →
↻ Miguel Alonso Jr. reposted
Lynn Cherny @arnicas.bsky.social

Google DeepMind article on 15 tears of Ai game research

Exploring new frontiers of AI and games research — Google DeepMind deepmind.google
AI Weekly's analysis →
  • Google DeepMind announced a research partnership with Fenris Creations, the studio behind EVE Online, to probe frontier AI capabilities inside the 20-plus-year-old space MMO.
  • The four capabilities in scope are named plainly: continual learning, memory, long-horizon planning, and complex multi-agent dynamics.
  • SIMA 2, powered by Gemini, now operates as an interactive companion across No Man's Sky, Valheim and Hydroneer, with an Aura Guidance system inside EVE.
Read full analysis →
View on Bluesky →
↻ Miguel Alonso Jr. reposted
Tim Kellogg @timkellogg.me

DSH Superagent an agent built on top of DeepSeek Harness that’s meant to stay out of the way, do work only according to team rules (e.g. pickup new bugs) github.com/deepseek-ai/...

GitHub - deepseek-ai/dsh-superagent-prototype: Towards a More Autonomous Agent System github.com
AI Weekly's analysis →
  • DeepSeek's Harness team released dsh-superagent-prototype on October 31, 2026 under MIT license, framed as research code and not an official DeepSeek product.
  • The design is a two-layer multi-agent split: one resident Router talks to people and dispatches work, isolated Workers run tasks in parallel.
  • The stated goal is to minimize human attention each autonomous action costs, with much of the engineering reportedly performed by coding agents themselves.
Read full analysis →
View on Bluesky →
↻ Miguel Alonso Jr. reposted
Tim Kellogg @timkellogg.me

I made a skill for Claude Code/Codex/etc. for selecting models based on Artificial Analysis reported benchmarks & costs I figure it'll be useful for letting models setup subagents well, based on the actual task at hand github.com/tkellogg/mod...

GitHub - tkellogg/model-selection: A skill for model selection github.com
AI Weekly's analysis →
  • The skill queries Artificial Analysis' API so subagents can rank available models by benchmark scores and cost before picking one for a task.
  • API calls are cached on a daily basis to stay under Artificial Analysis' free tier limit of 100 requests per day.
  • Install is a git clone into ~/.claude/skills, with an optional symlink into ~/.codex/skills for people running two agents.
Read full analysis →
View on Bluesky →
↻ Miguel Alonso Jr. reposted
Sung Kim @sungkim.bsky.social

GLM-5.3 is now open-weight. Their most capable model for agentic coding and cyber defense is now available to download, run, and customize. Weights: huggingface.co/zai-org/GLM-5.3 Tech blog: z.ai/blog/glm-5.3

z.ai
AI Weekly's analysis →
  • GLM-5.3 launched on August 14, 2026, keeping GLM-5.2's base and lifting Terminal-Bench 3.0 from 4.6 to 28.3 through post-training alone.
  • On CyberGym vulnerability discovery, GLM-5.3 scored 84.5%, slightly ahead of Claude Mythos 5 at 83.8% and GPT-5.6 Sol at 83.6%.
  • Z.ai says the model surfaced 2,436 vulnerabilities across 269 open-source projects, 1,097 rated critical or high, and delayed weights about two weeks.
Read full analysis →
View on Bluesky →

In Miguel Alonso Jr.'s orbit

Center = Miguel Alonso Jr.. Left = members they follow (green edges). Right = members who follow them (blue edges). Top = mutual follows (orange edges, slightly larger). Drag any node to reposition; click to open that profile.

Are you Miguel Alonso Jr.? Show it.

Add the Who’s Who of AI badge to your site or bio. It links back to this profile.

Listed in AI Weekly's Who's Who of AI

Markdown: [![Listed in AI Weekly's Who's Who of AI](https://aiweekly.co/modules/custom/aiweekly_whoswho/images/whoswho-badge.svg)](https://aiweekly.co/whos-who/person/miguelalonsojr-bsky-social)