self.md radar — 2026-05-06
Yesterday’s weekly issue was about the current map: memory is past, state is now, and a world model is what the system expects after it acts.
Today’s radar is the bill for that map. SubQ is trying to keep the whole repo in the room. Airbyte is pre-indexing SaaS mess so agents stop wandering. ProgramBench is the useful insult: even with a binary and docs, agents mostly cannot rebuild a real program from behavior alone.
1. SubQ put the whole-repo bet on attention math
sources:
what happened: Subquadratic launched SubQ 1M-Preview, calling it the first LLM built on a fully subquadratic architecture, with a 12M-token research result and a 1M-token product surface. Its SSA note says attention gets routed to content-dependent relevant positions instead of dense all-pairs attention; the company reports almost 1,000x less attention compute at 12M tokens and 52.2x faster input processing than dense attention at 1M tokens on B200s. The product page lists 150 tokens/sec, pricing at 1/5 of other leading LLMs, and benchmark rows for SWE-Bench Verified 81.8%, RULER @128K 95.0%, and MRCR v2 65.9%. The eyebrow stays up: the technical note says the benchmarks are third-party verified, but the full model card is still coming.
why this matters: If this holds, agent design changes from “which chunks should we retrieve?” to “who gets custody of the giant state surface?” This is the world-building point from yesterday’s weekly in less romantic clothes: the agent needs a current map of repo, history, permissions, and running state, and someone has to own it between steps.
2. Airbyte made context plumbing measurable
sources:
what happened: Airbyte launched Airbyte Agents as a data layer for agents working across operational systems like Slack, Salesforce, and Linear. The launch post from Michel Tricot says the trigger was a migrated agent trace that took 47 steps to answer “which customers are at risk of leaving this quarter?” and still answered wrong. Airbyte’s answer is a Context Store, populated by its replication connectors, exposed through MCP, a Python SDK, an API, and a web app. Its public benchmark repo compares Airbyte Agent MCP against vendor or native MCP servers for Linear, Slack, Gong, Salesforce, and Zendesk; the README says the benchmark measures tool-call token usage, not model quality, and the HN post reports up to 80% fewer tokens for Gong, 90% for Zendesk, 75% for Linear, and 16% for Salesforce.
why this matters: This moves the agent tax from vibes into a billable unit. If the agent is burning tokens to discover object IDs, map customers, and walk API pagination, the smarter component may be the dull index that stops it from wandering.
3. ProgramBench kept every frontier agent at zero
sources:
what happened: ProgramBench, from a Meta Superintelligence Labs, Stanford, and Harvard team, asks agents to rebuild real programs from a compiled binary and documentation. The benchmark has 200 tasks and uses mini-SWE-agent; agents run in containers with no internet access, execute-only permissions on the binary, and no decompilation tools. The public leaderboard is brutal in a clean way: Claude Opus 4.7, Claude Opus 4.6, Claude Sonnet 4.6, GPT 5.4, Gemini 3.1 Pro, Gemini 3 Flash, Haiku 4.5, GPT 5.4 mini, and GPT 5 mini all show 0% resolved. The near-miss column is barely warmer: Opus 4.7 at 3.0%, Opus 4.6 at 2.5%, Sonnet 4.6 at 1.0%, and the rest at 0.0%.
why this matters: This is the cold bucket beside the long-context champagne. More context can carry more state, but behavior-to-implementation still asks for architecture, taste, and a lot of negative space the tests do not hand you. The agent can run the binary and still not know the program.
supporting links
- world models for personal agents — the vocabulary behind today’s infrastructure story: memory is past, state is now, prediction is the next-state check.
- Bleeding Llama — Cyera’s Ollama memory-leak writeup is the custody footnote: local context servers still need boring network boundaries.
- Gemma 4 MTP — Google’s multi-token prediction drafters are another inference-cost lever, with the blog claiming up to 3x faster Gemma 4 inference.
- agent-browser — Vercel’s browser automation CLI is the small repo-gift version of the same story: agents want tools with handles, not screenshots and prayers.
left on the table
- DeepSeek V4 Pro on FoodTruck Bench stayed out because ProgramBench already owned the benchmark lane, and the source was mostly a Reddit/image claim.
- Hyundai and Boston Dynamics robots had a big number but weak primary grounding for a main slot.
- OpenCLI was an exact recent repeat from 2026-04-29.
- Vellum assistant already appeared as a supporting link yesterday.