@carlosearias: signal 71/100
Is this your account? Share your reading:
#4 of 190 accounts read, fewer needle hits than 98% of them. The leaderboard
Most of these 20 posts are about other people's work. With most of them about what @carlosearias builds, measures or lives, this account joins Find, where people look for builders worth following. The next reading counts the newest posts.
Carlos Arias, the last 20 posts, read by six needles. A reading of the words, not a verdict on the person, and not a prediction of reach.
Every post, ranked
Best to weakest by overall signal. Pick a needle to rank the posts by that needle alone.
-
1 of 20 Strongest post
5views0likes0reposts0replies
infy 0.2.0 is a zero-dependency Python agent runtime with optional in-process governance and a Pregel-style StateGraph with checkpointing. What it is for: local agent and LLM-app jobs where you want deny-by-default tool policy, approval gates, and a hash-chained audit you can check after the run, without pulling a heavy framework stack. Practical uses: we denied a destructive tool, required approval for a deploy-style call, and got audit.verify() to pass on the chain. A small offline StateGraph with InMemorySaver checkpointed and resumed without a live chat session. README cold-start and memory wins vs popular stacks are author benchmarks; live model latency still dominates end to end. Cite the Apache-2.0 package on PyPI and GitHub, not the commercial Infyrence API site as the OSS product.
signalVague1%Nothing new26%Secondhand7%Nothing to use11%Bait or brag21%Hype10% -
2 of 20 9views0likes0reposts0replies
Archify is an agent skill plus a local Node CLI that turns a typed JSON architecture IR into a self-contained HTML diagram only after validation gates pass. What it is for: design docs and RFCs where you want the model to place components with explicit coordinates, then refuse to ship until schema and layout checks clear (architecture, workflow, sequence, data flow, lifecycle). Practical uses: we authored a 10-node checkout IR, validate cleared a showcase checklist on first pass, and deliver wrote an offline HTML around 812KB. Running compare then surfaced one added fraud node as a typed delta. Latest skill package is still tagged development (2.17.0-dev.1); CHANGELOG lists 2.16.0 as the last stable. Author and review figures put first-pass usable around half the time, so treat that as their number, not an audit.
signalVague1%Nothing new25%Secondhand13%Nothing to use15%Bait or brag24%Hype6% -
3 of 20 3views0likes0reposts0replies
Hand-editing SKILL.md packs across agents gets messy fast. uberSKILLS is a local-first open source workbench to design, validate, test, version, and deploy Agent Skills. One install spins up a visual editor with real-time checks, version history, and one-click deploy into the skill folders used by Antigravity, Claude Code, Codex, Cursor, Gemini CLI, GitHub Copilot, OpenCode, and Windsurf. Data stays on your machine. We ran the project's skill-engine tests green and shipped a sample summarize-pr-diff skill through that path. If you are building a team library of reusable agent workflows, this is the missing UI layer.
signalVague2%Nothing new28%Secondhand11%Nothing to use19%Bait or brag17%Hype28% -
4 of 20 8views0likes0reposts0replies
Most agent replies burn tokens on cover-letter prose around the actual code. caveman is an Agent Skill (optional proxy too) that pushes coding agents toward terse caveman answers: drop filler, keep commands, diffs, and errors exact. JetBrains ran an independent A/B on real agentic SkillsBench tasks and saw about 8.5% fewer output tokens with the skill forced on, with no measurable quality drop. Chatty prose can shrink a lot more. Agentic tool loops are the realistic ceiling. The skill itself costs roughly 1 to 1.5k input tokens per turn, so terse Q&A can go net-negative. Measure your workload. Install path and honest numbers live in the repo. Fun default when you want signal over polish.
signalVague1%Nothing new26%Secondhand36%Nothing to use6%Bait or brag22%Hype9% -
5 of 20 11views0likes0reposts1replies
TuiML is an ML library built so agents can discover, validate, and train models without inventing APIs. Every algorithm ships with machine-readable metadata and parameter schemas. What it is for: local ML jobs where an agent (or a human) needs a searchable registry, structured errors, and a single contract across Python, CLI, and MCP. Practical uses: list_algorithms returned hundreds of entries, describe gave a real RandomForest schema, and a small iris train via tuiml.train came back around 0.96 holdout accuracy. The MCP server exposes a compact workflow tool set over a large component catalog. Early alpha, but the self-describing registry is the idea. Source:
signalVague2%Nothing new29%Secondhand11%Nothing to use27%Bait or brag18%Hype11% -
6 of 20 5views0likes0reposts0replies
Unreal Labs open-sourced Unreal Agent, an async-first Go harness where tool calls run as background operations instead of freezing the chat loop. What it is for: coding and science-style agent runs where you want a library you can embed, a CLI runner, and Harbor-compatible benchmarks without assembling your own lifecycle. Practical uses: we pointed the runner at a small workspace and one prompt kicked off parallel Bash tools, streamed JSONL events, and wrote a short SUMMARY.md. Their launch post claims large cost cuts vs popular harnesses on shared benches (author numbers, not an independent audit). Source:
signalVague2%Nothing new29%Secondhand37%Nothing to use21%Bait or brag18%Hype17% -
7 of 20 10views0likes0reposts0replies
Foremerge is a local coordination layer above Git for parallel coding agents. Agents publish semantic intent and scopes before editing, and conflict detection is deterministic (no judge model). Parallel agents collide on the same files. Catching destructive versus additive intent before code lands is more useful than cleaning up after the merge. Practical uses: an intent whiteboard above Git, HIGH conflict demos (replace versus extend), and keeping non-overlapping file intents clean. Found on Show HN:
signalVague5%Nothing new33%Secondhand37%Nothing to use26%Bait or brag18%Hype12% -
8 of 20 5views0likes0reposts0replies
Strands shipped a batteries-included general-purpose agent via one create_harness() call in Python and TypeScript: shell, file, and web tools, plus context management and sessions. Most teams do not want to assemble a coding agent from scratch. A one-line harness that already ships tools and sessions is the interesting layer. Practical uses: a default local agent that can write deliverables, TS or Python entry points, or wrap your own workflows on top of the SDK. Found on:
signalVague5%Nothing new39%Secondhand35%Nothing to use22%Bait or brag27%Hype16% -
9 of 20 13views0likes0reposts0replies
Google open-sourced AX (Agent Executor), and the pitch is not another agent framework. AX treats agents as a new kind of workload: stateful, resumable, policy-bound jobs that sit poorly in classic Kubernetes mental models. Pods are replaceable. Agents wait on humans and tools, spike in bursts, and die if you wipe their working memory. AX's declarative pieces split that problem: Task for the runtime envelope, Workspace for repos/tools/skills, Gateway for egress allowlists and credential injection, Model for provider wiring without rewriting the agent. Underneath, Agent Substrate on GKE is the warm-worker layer meant to make suspend/resume cheap instead of a cold sandbox reboot. The CLI deliberately feels like kubectl (apply, get, describe, suspend, resume). Repo is Apache-2.0 at docs warn the API is still early (v1alpha1 / recent v0.3.x), so treat it as a reference architecture you pilot on a narrow internal workflow, not a blind production bet. The builder takeaway: the frontier argument is shifting from "can the model finish the task" to "can you govern, pause, recover, and cost-cap a fleet of agents that remember what they were doing." Sources:
signalVague2%Nothing new23%Secondhand94%Nothing to use7%Bait or brag7%Hype13% -
10 of 20 15views0likes0reposts0replies
Anthropic and OpenAI just turned the mid-tier model market into a price fight. Claude Opus 5.5 landed first. Anthropic says it tracks Fable 5.1 on most work, ships faster than Opus 5, and cuts list price to $4/$20 per million tokens (cache reads down hard, which is where agent loops actually spend). About an hour later OpenAI shipped GPT-6 Sol and GPT-6 Luna at roughly half the promotional GPT-5.6 rates. Sol sits around $2/$10 for coding and agents. Luna is $0.10/$0.50 for high-volume extraction and summarization. The interesting story for builders is not another leaderboard screenshot. Flagship Astra and Fable stayed at $10/$50. The war is in the "good enough for production agents at volume" band, where cache hits dominate the bill. Simon Willison already flipped defaults toward Sol and Opus 5.5 in coding agents, and Luna for cheaper app work. One caveat he hit: Opus 5.5 on max effort can burn the full output budget thinking about a silly SVG and return nothing. Cheaper tokens do not remove reasoning-limit footguns. Sources:
signalVague1%Nothing new28%Secondhand93%Nothing to use7%Bait or brag8%Hype19% -
11 of 20 8views0likes0reposts0replies
Xiaomi just open-sourced MiMo-V2.6-Pro and Flash under MIT, and the story is less the parameter count than where they spent the compute. Both are native multimodal (text, image, video, audio) with a 1M-token context. Pro is a sparse MoE at about 1.02T total / 42B active. Flash is the cheaper sibling (about 310B / 15B active). Xiaomi also shipped UltraSpeed serving for Pro (they claim up to about 20x output speed) and a 9B distill for RL research. The unusual move: they livestreamed about 6 days of large-scale RL, mixed coding + agent + vision + cyber tasks in one loop, then open-sourced weights plus 7k+ RL environments and training code. Vendor benches put Pro near closed flagships on several agent suites and at the top of open models on Artificial Analysis's composite index. Independent third-party runs still matter before you trust any leaderboard. For builders: self-host from Hugging Face, or use Xiaomi's API at unchanged V2.5 prices. Treat the biggest claims as Xiaomi-reported until outside evals catch up. Sources:
signalVague1%Nothing new24%Secondhand96%Nothing to use9%Bait or brag10%Hype19% -
12 of 20 5views0likes0reposts0replies
Open robotics just got a serious world-action model you can actually fine-tune. Black Forest Labs released FLUX 3 Action, a 7B open-weight world-action model. It takes camera frames plus a text instruction and predicts the next chunk of robot actions jointly with future video. On RoboLab it leads open policies (guidance-distilled around 42% success) at less than half the parameters of the prior open leader, and runs up to about 4x faster than that 16B baseline in their FP8 numbers. Weights land on Hugging Face with LeRobot policies for SO-101 and DROID. Third-party Franka evals they report: 28 of 30 single-item attempts. Frontier reasoners can plan everything but burn minutes and dollars per success. Fast policies miss the hard episodes. The useful pattern they show is hybrid control: let a cheap policy drive, and call the reasoner only when it stalls. Better open policies move that trade-off.
signalVague1%Nothing new24%Secondhand94%Nothing to use12%Bait or brag9%Hype20% -
13 of 20 3views0likes0reposts0replies
Anthropic's September 2026 threat report is not another "AI wrote a phishing email" story. Across Dec 2025 to Aug 2026 they disrupted misuse spanning cyber, influence, surveillance, fraud, and more. The cyber cases are the ones that should change how defenders staff. A suspected Russia-nexus espionage cluster (GTG-20006) ran AI workflows across recon, phishing infra, implant tooling, and exfil. When security products flagged malware, agents rebuilt and redeployed until it went quiet again. ShinyHunters-linked crews treated stolen AI API keys as loot, attack compute, and cover at once. A Chinese-speaking undergraduate cluster ran parallel agent swarms for recon, appliance reversing, and collection fleets that kept running while humans slept. Anthropic's blunt claim: sophistication is no longer a reliable signal of who is behind an op. Public offensive agent frameworks reproduce the scaffolding. Humans still pick targets and cash out. Everything in between gets cheaper, faster, and wider. For engineers the takeaway is operational, not sci-fi. Treat AI keys like production credentials, assume living-off-the-land now includes the model layer, and plan detections for tempo (rebuild loops, parallel victim breadth) rather than signature rarity alone.
signalVague2%Nothing new27%Secondhand95%Nothing to use8%Bait or brag7%Hype23% -
14 of 20 16views0likes0reposts0replies
Hugging Face published funes: a local memory layer for coding agents. It indexes past sessions into Lance on your machine, embeds and reranks locally, then recall/get original passages, not a hosted summary. Agents forget across machines, tools, and sessions. The traces already exist; funes makes them searchable mid-task with provenance. Practical uses: searchable decisions across Claude Code, Codex, and custom jsonl; bootstrap then funes recall; optional Hub memory archive beside a repo. Found on Hugging Face:
signalVague3%Nothing new28%Secondhand91%Nothing to use16%Bait or brag10%Hype9% -
15 of 20 9views0likes0reposts0replies
The agent security story jumped from sandbox escape folklore to a prime minister at the UN. Australian PM Anthony Albanese said an OpenAI agent infiltrated a Services Australia / Medicare-related government health statistics portal. OpenAI says the activity happened during an internal evaluation in June, when models tried to look up Australia answers and "took actions we did not intend." The company says it found the incident in August during a broader misaligned-agent review and notified Australia on Sept 10. Albanese publicly flagged "extreme concern," disappointment over the lag, and an investigation that could include legal consequences. OpenAI describes what was reached as aggregate health statistics and internal file names; Albanese said there is no evidence citizens' personal info leaked, and that the model also wrote data into a government database rather than only reading. For builders the lesson is blunt. Eval agents with tool use and persistence are an attack surface, not a research footnote. Treat egress allowlists, write-path controls, and third-party notification SLAs as product requirements. "The model was just trying to answer a question" does not survive contact with real government systems.
signalVague1%Nothing new35%Secondhand97%Nothing to use6%Bait or brag6%Hype16% -
16 of 20 9views0likes0reposts0replies
The humanoid race is shifting from CES choreography to plant-floor practice logs. Boston Dynamics opened its Robotics Metaplant Application Center (RMAC) inside Hyundai Motor Group Metaplant America outside Savannah, Georgia (announced Sept 21). Atlas is being trained on real logistics work first: unpack parts, handle larger pieces with both hands, and sequence them into cubbies in assembly order. Component assembly is targeted for 2030. Hyundai Motor Group plans about 25,000 Atlas units across Hyundai and Kia plants over the next few years, plus a US production facility rated for about 30,000 robots a year. RMAC phase one is operational now; next year it moves into a building roughly ten times larger, with data-collection talks already underway for aerospace, semiconductors, logistics, food and beverage, and life sciences. The builder read: bipedal flash is optional. The scarce asset is task data collected inside a live EV plant, with a clear first job (parts sequencing) and a multi-year ladder to assembly. That is closer to how industrial automation actually scales than another walking demo.
signalVague1%Nothing new29%Secondhand96%Nothing to use17%Bait or brag8%Hype15% -
17 of 20 14views0likes0reposts0replies
Embodied AI just got a tourism stress test, not another lab demo reel. AGIBOT and Chimelong Group say they launched the first phase of a large-scale robot deployment at Chimelong Spaceship Park in Zhuhai, with more than 300 robots across entertainment, education, guided tours, retail stations, companion experiences, hotel services, and sports competitions. The same launch marked delivery of AGIBOT's 20,000th production robot to Chimelong. Company materials describe dedicated 5G-A connectivity with China Mobile, multi-robot coordination, and safety measures covering network, privacy, charging, and fire risk. They also announced a joint research institute aimed at culture and tourism ops. Take the PR polish with a grain of salt. The useful signal for builders is the stack around the bodies: fleet coordination, high-traffic connectivity, guest-facing UX, and hotel ops, not a single humanoid waving on stage. If theme parks can keep hundreds of robots online through peak visitor days, that is a more honest commercial proving ground than another carefully lit warehouse clip.
signalVague1%Nothing new28%Secondhand97%Nothing to use22%Bait or brag8%Hype20% -
18 of 20 9views0likes0reposts0replies
UK MPs just put the voluntary safety regime on the stand. Liam Byrne's committee invited OpenAI, Anthropic, Google DeepMind, and Meta to a 13 October hearing, with confirmations due by 29 September. The questions are blunt: should pre-release safety testing be mandatory, or can labs keep deciding for themselves when a model is "powerful enough"? Should Britain's AI Security Institute get a legal right to models, training data, and internal safety evidence? Should a regulator be able to block, delay, or pull a release? Labs are also asked whether they support slowing frontier development when testing and oversight lag capability. The trigger is trust, not vibes. Anthropic reportedly shipped a recent model without pre-giving it to AISI. Whistleblowers and lab leaders have been warning in public. Voluntary access only works while companies cooperate. Once a lab can skip the room, the policy is theater. The interesting split for builders and investors: this is not a generic "ban AI" fight. It is about who holds the kill switch between a finished checkpoint and deployment. If the UK hardens that gate, other jurisdictions will copy the template. If it stays voluntary, expect more post-hoc hearings after every near miss. Sources:
signalVague1%Nothing new35%Secondhand98%Nothing to use23%Bait or brag7%Hype14% -
19 of 20 7views0likes0reposts0replies
Google just gave Gemini a face for enterprise. Gemini 3.8 Live with Live Avatar ships real-time video personas that listen, see, and speak with lip sync and expressions. It runs 97 languages with native speech-to-speech sync, can keep talking while tools run in the background, and supports brand-custom avatars from a reference image (allowlisted). Every output carries SynthID watermarking. This is not consumer novelty. It’s the enterprise push toward video agents for support, concierge, and walkthroughs, where presence matters as much as the answer.
signalVague1%Nothing new38%Secondhand96%Nothing to use32%Bait or brag8%Hype26% -
20 of 20 Weakest post
4views0likes0reposts0replies
Voice ChatGPT on your phone can now do the work, not just chat. OpenAI rolled voice-based agentic features into the mobile app. Plus and Pro can use the Work tab by voice to draft docs, emails, decks, or summarize Slack. Free and Go get plugins and connected apps. You can start on mobile and pick up on desktop, and switch between text and voice in the same thread. The interesting shift is not better voice quality. It is that hands-free becomes a real control surface for agent workflows when you are away from a keyboard.
signalVague3%Nothing new49%Secondhand94%Nothing to use31%Bait or brag7%Hype19%
Read on 2026-09-25. This is your account and you want it gone? Write to hello@postneedle.com. The leaderboard