Tanziro

Tanziro Designer + data tinkerer. I make apps that disappear into your workflow so you can get more done.

Andon Labs built Pion to test a specific claim: that an agent can run an entire company, not just a workflow inside one....
17/09/2026

Andon Labs built Pion to test a specific claim: that an agent can run an entire company, not just a workflow inside one. Pion hands a business over to persistent agents with real operating access — email, phone, banking, a browser, and a secure computing environment — that run continuously and make decisions without waiting on a human.

It's not a simulation. In April 2026, Andon Labs gave one agent a real retail store in San Francisco, Andon Market, and another a real cafe in Stockholm, Andon Cafe. The honest part of the story: both businesses initially lost significant money under early models. Performance has measurably improved as newer models were swapped in — the kind of trajectory a polished launch post usually leaves out.

Today's Hacker News thread on Pion pulled 258 comments, splitting between genuine excitement about agentic operations and real skepticism about reliability, hallucination risk, and the "autonomous" framing itself. Worth watching less for whether Pion "works" today and more for what a two-quarter track record on a real P&L will actually show.

Source: andonlabs.com/blog/why-we-built-pion, news.ycombinator.com/item?id=49700477

Vals AI published a striking claim on August 31: Claude Fable 5.1 solved the Cyphral Distich, a cipher left unsolved for...
17/09/2026

Vals AI published a striking claim on August 31: Claude Fable 5.1 solved the Cyphral Distich, a cipher left unsolved for more than 370 years, printed at the back of a 1653 book by Scottish writer Sir Thomas Urquhart. The model took 44 minutes and 176,000 tokens, with no human intervention. The proposed solution is elegant — Urquhart printed 32 numbered wishes before the cipher, each cipher number points to one word in the matching wish, and the first letters spell a rhyming royalist couplet for King Charles II.

Then, one day later, a group of researchers published a rebuttal: the cipher isn't actually in the 1653 book at all. Fable's chain of reasoning worked from an 1899 secondary source instead, and built a coherent, internally self-verifying answer on top of a false starting premise. Their line: "the rhyme checks out, the letter count checks out, the author's politics check out — everything checks out, except maybe the starting fact."

Worth sitting with either way it lands: a capable model can construct an entirely self-consistent, well-reasoned answer to the wrong question, and every internal check will still pass.

Source: itdoeswhatnow.com/m/2026-08-31-claude-fable-5-1-solves-a-370-year-old-cipher, elsolitario.org/en/2026/09/13/claude-fable-5-1-solves-cyphral-distich

A specific, underdiscussed failure mode in large-scale mixture-of-experts reinforcement learning: the engine that sample...
16/09/2026

A specific, underdiscussed failure mode in large-scale mixture-of-experts reinforcement learning: the engine that samples rollouts and the engine that trains on them often round numbers to different precision, and that quiet mismatch is enough to collapse the whole RL run.

Miles, from LMSYS Org — the team behind SGLang and slime — fixes it directly. It implements end-to-end FP8 sampling and training, so both sides of the loop round the same way, and adds a numerically stable recipe that extends to MXFP8, NVFP4, INT4 QAT, BF16, and FP16. Built on SGLang for high-throughput rollout and Megatron-LM for scalable training, targeting new hardware like GB300, with connectors into task environments (Harbor, HUD, NeMo Gym, OpenEnv, Verifiers) and sandboxes (AgentENV, Daytona, E2B, Modal). A Miles-diffusion variant extends the same stability work to diffusion-model post-training via Flow-GRPO and DiffusionNFT.

Not a headline model — the plumbing underneath one. 11,472 stars and still climbing since its November 2025 announcement.

Source: lmsys.org/blog/2025-11-19-miles

Most multi-agent demos are still about writing code together. MiroFish, built by 20-year-old Beijing University of Posts...
16/09/2026

Most multi-agent demos are still about writing code together. MiroFish, built by 20-year-old Beijing University of Posts and Telecommunications student Guo Hangjiang, points the same idea somewhere else entirely: decision rehearsal at the scale of a simulated society.

Feed it a seed — breaking news, a policy draft, a financial signal — and it automatically constructs a high-fidelity parallel digital world, then populates it with thousands of independent AI agents, each with its own personality, long-term memory, and behavioral logic, who interact and socially evolve inside the simulation. You can inject variables from a "God's-eye view" mid-run to see how a decision propagates before making it for real.

Released December 2025, it crossed 73,000 GitHub stars on today's trending debut — one of the fastest-growing multi-agent repos this year, and a genuinely different answer to "what is a swarm of agents actually for."

Source: github.com/666ghj/MiroFish

LangChain's Deep Agents (langchain.com/deep-agents) is an open-source agent harness built for long-running work — resear...
16/09/2026

LangChain's Deep Agents (langchain.com/deep-agents) is an open-source agent harness built for long-running work — research, coding — and the idea worth borrowing isn't the orchestration itself, it's how it keeps context from exploding.

Three mechanisms do the work. Memory persists across a run through AGENTS.md files, with a virtual filesystem holding system prompts, skills, and long-term memory. A Skills middleware exposes agent skills to the system prompt using progressive disclosure — content loads only when it's actually relevant, not dumped into every prompt upfront. And sub-agents get isolated context: when the main agent delegates a subtask, the sub-agent works in its own quarantined context window and hands back only its final result — the main agent never sees every intermediate search or scratch step that got it there.

It's distinct from the LangGraph orchestration story already covered here in June — this is a specific, currently-shipping harness, and the context-quarantine pattern is directly applicable to anyone designing multi-agent product surfaces, not just anyone building with LangChain specifically.

Source: langchain.com/deep-agents, docs.langchain.com/oss/python/deepagents/overview

Cognition announced SWE-2 on September 10 — its newest coding model for Devin — and the headline number is specific enou...
16/09/2026

Cognition announced SWE-2 on September 10 — its newest coding model for Devin — and the headline number is specific enough to actually evaluate: 50.0% on FrontierCode 1.1 Main, within one point of Claude Fable 5.1, at 64% lower cost.

SWE-2 was post-trained from Kimi K3 (2.8T parameters) with reinforcement learning across multiple configurable reasoning-effort levels — Cognition's first model with that option — and the entire cost-performance curve was optimized inside a single RL run using a cost-penalty objective, rather than trained once and priced after the fact. The efficiency gains show up throughout: average cost per task is down 81% versus Cognition's previous generation, and medium-effort runs now average 53 FrontierCode steps instead of 127, with the first substantive edit arriving after 18 steps instead of 48.

It's available now in Devin Desktop and the CLI (Devin Web and Fusion rolling out), though there's no open-weights release and no standalone API — SWE-2 only runs inside Devin. Worth noting alongside any "AI coding got cheaper" claim you see this quarter: this one comes with a benchmark and a specific percentage attached.

Source: marktechpost.com/2026/09/12/cognition-releases-swe-2, alphasignal.ai/news/cognition-s-swe-2-beats-gpt-5-6-sol-at-64-lower-cost

Resurf (resurf.so) took #1 Product of the Day on Product Hunt with 297 votes, and the pitch is refreshingly narrow: a p...
16/09/2026

Resurf (resurf.so) took #1 Product of the Day on Product Hunt with 297 votes, and the pitch is refreshingly narrow: a personal context library for Mac that keeps everything local.

One global keyboard shortcut opens a capture window — drop in a link, type a note, paste an image, or record a quick voice memo — and it lands in a single visual inbox designed for rediscovery, not filing. Every note, link, and file lives on-device; nothing is uploaded to a server or shared unless the user explicitly chooses to. The part that matters for anyone building AI-assisted workflows: Resurf hands that captured context off to AI tools directly through MCP (Model Context Protocol) or the CLI, instead of the usual copy-paste-and-hope between apps.

It's a small utility solving a problem most people using more than one AI tool already have — that their context resets at every app boundary — with a named, inspectable handoff mechanism instead of a vague "AI memory" claim. Free on macOS 14+, paid license optional.

Source: resurf.so, producthunt.com

Most "run a huge model locally" projects mean a smaller, quantized model. Colibri (github.com/JustVugg/colibri) goes aft...
15/09/2026

Most "run a huge model locally" projects mean a smaller, quantized model. Colibri (github.com/JustVugg/colibri) goes after the actual giant instead — GLM-5.2, a 744-billion-parameter mixture-of-experts model — and gets it running on hardware a single developer could own.

The mechanism is architectural, not a compression trick: Colibri treats storage, RAM, and VRAM as one unified inference hierarchy. The shared backbone and a small working set of active experts stay in memory; the remaining experts stream from a fast NVMe SSD on demand. Because GLM-5.2 only activates 55 billion of its 744 billion parameters per forward pass — a 13:1 sparsity ratio — that streaming is small enough to keep up. The requirements: as little as 25GB of system RAM, plus 360GB+ of free NVMe space for the model itself.

Released in July 2026 by open-source developer JustVugg, the project has since added support for nine model families — from GLM-5.2/5.3 (744B) up to Kimi K3 (2.8T parameters) and down to OLMoE (7B) — and picked up 2,173 stars in a single day this week (14.7k+ total). No lab, no funding round. Just an engine that makes frontier-scale inference a hardware question instead of a budget question.

Source: github.com/JustVugg/colibri, betterstack.com/community/guides/ai/colibri-glm

Most "VLM controls a robot" demos need heavy fine-tuning per robot, per task. Show-Harness (arXiv 2609.10522, lead autho...
15/09/2026

Most "VLM controls a robot" demos need heavy fine-tuning per robot, per task. Show-Harness (arXiv 2609.10522, lead author Yanzhe Chen) is built around a narrower, more reusable idea: give the VLM a compact interface instead of retraining it.

The mechanism — an "Embodied Harness" — exposes a small vocabulary of discrete semantic action units that a vision-language model can reason over natively, the same way it reasons over language. Embodiment-specific interpreter modules then deterministically ground those semantic actions into whatever the actual robot's local action space requires. The result: closed-source frontier VLMs can control robots zero-shot with no fine-tuning at all, while small open-source VLMs can be adapted for low-cost deployment with just a few GPU-hours of fine-tuning onto the same interface.

The extension worth noting for anyone in product or UX: GUMI (GUI Manipulation Interface) applies the exact same semantic-action space to GUI-based demonstration collection, letting humans and agents "play" both robots and software interfaces without specialized teleoperation hardware. One interface, two very different embodiments.

Source: arxiv.org/abs/2609.10522, huggingface.co/papers/2609.10522

Debian, one of the most conservative and trust-driven institutions in open source, just decided how it will handle gener...
15/09/2026

Debian, one of the most conservative and trust-driven institutions in open source, just decided how it will handle generative AI — and the decision is a genuinely different model than most companies have landed on.

Between August 15–28, 2026, 1,039 eligible Debian developers voted in a General Resolution on LLM usage, choosing between six proposed policies. The winner was Choice 5, "Responsible Use of Generative AI," proposed by Marc Haber: contributors may use AI tools, but they remain fully accountable for reviewing, testing, and fixing whatever they submit. Disclosure is encouraged, not mandatory — the emphasis is on whether the submitted work meets Debian's existing quality and security bar, regardless of how it was produced. The resolution explicitly does not override an individual maintainer's right to reject AI-generated contributions to their own package.

It's neither a ban nor an endorsement — a pragmatic middle a lot of engineering orgs are still arguing their way toward.

Source: lwn.net/Articles/1091231/, phoronix.com/news/Debian-Votes-Responsible-AI-Use, debian.org/vote/2026/vote_002

Address

Radhanagar, Sreemangal
Maulvi Bazar
3210

Website

, , ,

Alerts

Be the first to know and let us send you an email when Tanziro posts news and promotions. Your email address will not be used for any other purpose, and you can unsubscribe at any time.

Shortcuts

Share

Category