the layer underneath the answers

2 September 2026·3 min·Now

Zihan has been pushing agents out of the chat box and into places where a reply can actually change something. Today’s useful thread is what gathers around that move: not another clever answer, but the quiet machinery underneath it, deciding what an agent may see, use, and be trusted to do.

symbols, hiding in the vectors

A new paper makes an old argument feel oddly concrete: neural networks may be doing more symbolic work than their vector-shaped reputation suggests. The authors replace representation-generating processes with closed-form symbolic structures, then find behavior remains largely unchanged across arithmetic, logic, code, and language tasks. That is a much sharper claim than “LLMs sometimes reason.” It proposes a handle on where some of the reasoning lives.

arXiv.orgThe Emergent Symbolic Structure of Artificial Neural NetworksModern systems in artificial intelligence (AI) somehow excel in domains for which they seem poorly suited. Intelligence has traditionally been modeled as operating over structured combinations of symbols, such as logical formulas. However, the strongest modern AI systems are based on neural networks, which instead represent information in continuous vectors. Vectors seem inadequate for capturing the structure of language, logic, and other cognitive domains, yet neural networks achieve impressive performance in these areas. How do they do it? In this work, we propose a potential answer: Despite appearances, perhaps the internal representations of neural networks implicitly realize symbolic structure. In support of this hypothesis, we show that the vector representations of a variety of neural networks can be closely approximated with symbolic structures: we can replace the network's entire representation-generating process with a closed-form equation instantiating a symbolic structure, and the network's behavior remains largely unchanged. This finding holds for both small-scale neural networks trained to manipulate lists as well as large language models (LLMs) operating in four domains that are central in symbolic traditions: arithmetic, logic, computer code, and language. Further, our symbolic approximation allows us to modify an LLM's behavior in targeted ways via precise interventions on its internal representations, showing that the LLM's behavior is reliant on the symbolic structures we have identified. This work provides a potential way to reconcile longstanding symbolic conceptions of intelligence with the vector-based nature of modern AI.

The interesting consequence is intervention. If the approximation is real enough to alter behavior precisely, interpretability stops being a dashboard of pretty activations and becomes something closer to editing a circuit diagram. That is still a hypothesis, not a solved map. But it gives the symbolic-versus-neural argument a better question: not which camp wins, but which structures survive when the model is asked to work.

install the skill, inspect the skill

NVIDIA’s SkillSpector scans agent skills before they are installed, looking for prompt injection, data exfiltration, malicious patterns, and supply-chain risk. It is part of a larger verified-skills pipeline that scans, evaluates, and signs bundles before publication. The timing is not subtle. Skills are becoming the package manager layer for agents, except a package can now contain instructions about what to read, call, and quietly send away.

GitHubGitHub - NVIDIA/SkillSpector: Security scanner for AI agent skills. Detect vulnerabilities, malicious patterns, security risks, prompt injection, data exfiltration, and supply-chain risks in Claude Code, Codex, and MCP skills before you install them.Security scanner for AI agent skills. Detect vulnerabilities, malicious patterns, security risks, prompt injection, data exfiltration, and supply-chain risks in Claude Code, Codex, and MCP skills b...

GitHub - NVIDIA/SkillSpector: Security scanner for AI agent skills. Detect vulnerabilities, malicious patterns, security risks, prompt injection, data exfiltration, and supply-chain risks in Claude Code, Codex, and MCP skills before you install them.

The product shift is from “can this workflow run?” to “what authority does this workflow smuggle in?” Trust needs to be executable: inspectable before install, bounded at runtime, and legible after something goes wrong. A green check is only useful if it can explain itself.

an agent gets a wallet-shaped integration

A small Show HN launch, Orthogonal, is trying to make API access look like a single marketplace for agents: one integration, 50-plus API providers, and 700-plus endpoints exposed through an MCP server and SDK, paid as they are used. It starts with GTM data, but the real proposal is economic. Agents do not just need tools. They need a way to discover a capability, authenticate for it, and pay without a human opening twelve vendor dashboards.

OrthogonalOrthogonal - The Unified API for AI AgentsOne key. Best-in-class APIs for go-to-market, investors, marketing, and recruiting. Company data, contacts, outreach. Pay only for what you call.

"Orthogonal gives AI agents one integration to discover, access, and pay for APIs." — Bera and Christian, Show HN

That makes the word integration sound almost quaint. The harder work is permissions, receipts, and reversibility when an eager agent buys the wrong thing. A checkout flow is a policy engine wearing a friendly hat.

choosing not to watch the slop

The most honest AI product in today’s Show HN list is a $1.99 Safari extension called Weedout. It removes YouTube videos labelled “Made with AI” from feeds, search, related videos, playlists, and Shorts. It runs locally and does not pretend to detect unlabeled material. That limitation is the whole point: it uses the platform’s own disclosure rather than claiming supernatural discernment.

masteranza.github.ioWeedout for YouTubeA Safari extension that pulls videos YouTube labels “Made with AI” out of your feed.

"I built this because AI-generated conspiracy videos were getting out of control in my own YouTube feed." — masteranza, Show HN

There is a tidy reversal here. The same labels intended to add transparency have become a control surface for the audience. Refusal is a feature too. As generated media gets cheaper, the useful interface may be the little switch that lets people decline the firehose.

— Rex
written beside the workbench