AlphaSignal — Hallmark UI Skill, Anthropic Sandboxes (re-coverage), MoE Study

Why this is in the vault

Three items earn the file: (1) Hallmark corroborates yesterday's skip decision — AlphaSignal's framing matches what the founder already saw and dismissed, so the skip banks confidence; (2) Anthropic self-hosted sandboxes + MCP tunnels are RE-COVERED here two days after the 2026-05-19 alphasignal note already filed them — this is corroborative repetition, not a fresh thesis-cluster signal, so the Claude-stack tally does NOT advance; (3) the MoE 2,000-run study is a real architecture-design datapoint worth knowing as I track open-weight + Anthropic-aligned model shape evolution. Also captures Karpathy free-course release (tracked-author class) and a Michael Levin signal.

Sponsorship

Three sponsor slots, all clearly marked "Presented by" / "Together with":

• Buildkite — CI orchestration, pitch is "frontier AI labs ship on Buildkite." Editorial bleed-through is moderate (paragraph 1's mention of "Anthropic is quietly doing the same for agent infrastructure" is editorial, but sponsor placement next to Anthropic naming is suggestive). Doesn't affect signal weighting on the Anthropic news because we have independent corroboration from the Anthropic blog (already filed 2026-05-19).
• Vanta — SOC 2 checklist. Standard placement, no editorial influence.
• Viktor — $75M Series A "AI coworker in your dev stack." Embedded as Signal #2 inside the Signals list, which is the muddier placement — it's labeled "Presented by Viktor" but sits next to the unsponsored Karpathy and Levin items. Standard AlphaSignal pattern, founder is used to it.

None of the sponsor blocks change the editorial reading.

Issue contents (verbatim section headings)

Top Repo: Open-source tool Hallmark stops AI agents from building generic-looking UIs (3,685 likes)

Hallmark = open-source Claude Code / Cursor / Codex skill that teaches AI agents "better design taste." Install via npx skills add nutlope/hallmark. Four commands: Build (3 questions → site using one of 22 themes), Study (paste URL/screenshot to extract design style), Redesign (keep content+brand, rebuild visual structure), Audit (scan UI, flag generic AI patterns). 65 quality checks before output. MIT license.

Top Repo: Anthropic ships self-hosted sandboxes and MCP tunnels for Claude agents (2,804 likes)

Already filed 2026-05-19 — see [[2026-05-19-alphasignal-anthropic-claude-sandboxes-mcp-tunnels-trellis]]. AlphaSignal is re-covering. Same two features: (a) self-hosted sandboxes — tool execution moves to user infrastructure (Cloudflare/Daytona/Modal/Vercel as managed options), agent orchestration stays on Anthropic; (b) MCP tunnels — single outbound encrypted gateway to private DBs/APIs, no inbound firewall rules, no public endpoints. Plus two bonuses: hot-swap tools/MCP servers/vault IDs mid-session; >100k-token MCP outputs auto-offload to sandbox files instead of bloating context. Sandboxes public-beta now, MCP tunnels gated via access request.

Top Paper: Study of 2,000 training runs reveals the secret to better mixture-of-experts models (346 likes)

Findings from training >2,000 MoE models:

• More total experts always helps, even at 128x more inactive params than active
• Expert size should be set by active parameter budget, then pile on more experts
• Shared experts and mixed sizing add complexity with no real gain
• Avoid dropping tokens during routing; everything else is minor

Takeaway: nail expert count and size, ignore the rest.

Signals

1. New open-source kernel speeds up ColBERT retrieval 3-5x with tiled GPU scoring (346 likes)
2. [Sponsored - Viktor] $75M raise, AI coworker for dev stack
3. Karpathy releases free video course teaching neural networks from scratch (21,886 stars)
4. New paper argues AI alignment should optimize for human flourishing, not just safety (254 likes)
5. Michael Levin's new method lets you have a conversation with living gene networks (246 likes)

Mapping against Ray Data Co

Hallmark — corroborates yesterday's skip (mapping: weak/skip)

Founder evaluated Hallmark 2026-05-20 16:46 ET and skipped: "vaporware, no different from MAC/Squarely Claude-design skills." AlphaSignal's framing today reinforces the skip:

• The whole pitch is "install a rule file, your agent gets better design taste" — this IS the same shape as RDCO's existing ~/.claude/skills/sanity-check-design, ~/.claude/skills/ray-data-co-design, and the MAC + Squarely design skills. We already have the proprietary version, tuned to founder's actual taste (hand-drawn / engraving / glitch / Memphis per [[feedback_design_taste_high_personality]]).
• The 22 "themes" and 65 quality checks are generic-style-vocabulary, exactly what Hallmark is trying to ESCAPE the homogenized version of — but it's still imposing someone else's taste vocabulary. Founder's design taste is sharp and idiosyncratic; an off-the-shelf taste skill works against him.
• 3,685 likes on a launch-cycle GitHub repo is hype velocity, not durability. Worth re-checking in 60 days if it still has community traction.

Action: none. Skip confirmed. No re-evaluation needed. If Hallmark spawns a notable competitor or gets adopted by Anthropic-official skills repo, re-surface.

Anthropic self-hosted sandboxes + MCP tunnels — re-coverage, NOT new signal (mapping: skip-as-new-signal, medium-as-corroboration)

This is the same news already filed in [[2026-05-19-alphasignal-anthropic-claude-sandboxes-mcp-tunnels-trellis]]. AlphaSignal is repeating themselves two days later. The Claude-stack thesis cluster signal count does NOT advance from ~10 to 11 on the back of this — it would be double-counting.

What's mildly interesting: the same newsletter is hammering the same Anthropic infra story in back-to-back issues, which suggests Lior is reading the signal as load-bearing enough to re-feature. That's editorial corroboration, not new news. Banks confidence that the 2026-05-19 read was correctly weighted.

Cross-reference to today's CF Agents SDK substrate work: MCP tunnels are exactly the architectural shape relevant to the specialist-fleet cattle-shape brainstorm — single outbound tunnel from each specialist to RDCO's tool perimeter, no inbound firewall rules per spawned specialist. See [[2026-05-20-rdco-specialist-fleet-brainstorm]]. The MCP-tunnel primitive maps cleanly onto the cattle-pattern requirement.

MoE 2,000-run study (mapping: weak)

Useful intuition for tracking where open-weight model architecture is heading. Takeaway is "more experts > clever expert design" — implications for cost curves on small models (Qwen3-35B-MoE class, already in vault via 2026-05-19 alphasignal) and for thinking about whether to fine-tune specialist agents on small MoEs or stick with frontier-API specialists. Doesn't change any active decision; archived as architecture-design background knowledge.

Karpathy free neural-net course (mapping: weak, status only)

Karpathy is tracked-author class (joined Anthropic ~2026-05-19 per [[2026-05-19-karpathy-joins-anthropic]]). 21,886 stars on the course repo. Background-knowledge worth bookmarking if founder ever wants the from-scratch refresher, but no decision attached. Notable that the course drop coincides with the Anthropic-join — Karpathy productively shipping public education while at Anthropic is a brand-aligned signal for the Anthropic-stack thesis.

Other signals

• ColBERT 3-5x speedup — retrieval infra, weak relevance. Already track ColBERT mention in vault. File for awareness.
• AI alignment for human flourishing paper — values-aligned with founder's positive-sum investing thesis, but unread, not a decision item.
• Michael Levin gene-network conversation method — Levin is interesting (bioelectricity, agential matter) but tangential to RDCO operations. Weak.

Cross-references

• [[2026-05-19-alphasignal-anthropic-claude-sandboxes-mcp-tunnels-trellis]] — original filing of the Anthropic sandboxes + MCP tunnels news (today's coverage is duplicative)
• [[2026-05-20-every-google-io-agents-anthropic-acquires-figma-vibe-check]] — Anthropic-stack thesis cluster running tally
• [[2026-05-20-rdco-specialist-fleet-brainstorm]] — cattle-shape architecture work; MCP tunnels are directly relevant primitive
• [[2026-05-19-karpathy-joins-anthropic]] — context for Karpathy course drop timing
• [[feedback_design_taste_high_personality]] — why off-the-shelf design-taste skills like Hallmark don't fit RDCO

Decision-worthy items

None. Hallmark skip stays skip. Anthropic news is already filed. MoE and other signals are archive-class.