AlphaSignal — Mistral Large 4 ("Le Chonk") 1T-param open model, Claude lands natively in Google Workspace, Teleport's MCP-auth gap
Why this is in the vault
Two items worth keeping: Mistral's 1T-parameter open-weight model beating closed frontier models on aggregate benchmarks (continuing the open-weight-frontier-model trend already tracked from Reflection Beam and DeepSeek), and Claude shipping as a native Google Docs/Sheets/Slides sidebar — a direct product-surface overlap with how RDCO already works inside Google Workspace.
Mapping against Ray Data Co
Most concrete connection: the Teleport-sponsored item ("Why MCP Breaks Between Prototype and Production") reports 492 MCP servers sitting on the open internet with zero auth, and 67% of security leaders still relying on static credentials for AI — this is the exact failure mode feedback_mcp_install_security_review_default exists to prevent (security-review third-party MCP/plugin/skill installs before use, even unasked). It's sponsor content, not independent research, but the underlying stat (492 unauthenticated servers) is a useful external data point reinforcing why that policy stays a hard default rather than a judgment call.
Second connection: Claude's native Google Docs/Sheets/Slides integration (sidebar add-on via the Workspace Marketplace, or connector-based file-open-next-to-chat) lands in the same product direction as the 2026-09-17 issue's Claude Docs/Slides/Design merge already filed (2026-09-17-alphasignal-claude-docs-slides-design) — Anthropic keeps collapsing the app-switching RDCO currently does via its own Google Workspace skill and Canva connector into native in-chat tools. Worth a hands-on check of whether the Workspace-Marketplace sidebar changes anything about the existing docs-skill workflow, though nothing here forces an immediate change.
Weakest connection: Mistral Large 4 ("Le Chonk," 1T total / 49B active params, MoE, Apache-weighted open release end of October, 1M context, 160+ languages) is a watch-list open-weight entry, same shape as Reflection Beam and DeepSeek before it — no current RDCO workload needs a self-hosted frontier-class model, so this is tracked for the open-weight-economics trend, not actioned. Note a pricing inconsistency in the source itself: the subject line advertises "$1.18/M out" but the body text states "$1.36/M input and $4.18/M output tokens" — neither figure is independently verified here; flagging the discrepancy rather than picking one silently.
OpenAI's 722-manuscript math-research drop (Fields-Medal-advised, Lean-verified proofs) has no direct RDCO hook and is noted for completeness only.
Curation section
- Mistral Large 4 "Le Chonk" (Top News, 38,036 likes) — 1T total parameters, 49B active per token (MoE), API live now via Mistral Studio ($1.36/M in, $4.18/M out per body text; subject line says "$1.18/M out" — discrepancy not resolved), 1M token context (~750k words), native image input, 160+ languages, runs cybersecurity tasks closed models typically refuse. Open weights drop end of October; built entirely on Mistral's own EU GPU cluster, claimed strongest open-weights model from the US or Europe on aggregate benchmarks. Not deep-fetched (watch-list item, no current self-hosting need).
- Presented by Teleport — "Why MCP Breaks Between Prototype and Production": 492 MCP servers exposed on the open internet with zero auth; Teleport pitches certificate-based auth + mTLS; cites 67% of security leaders still on static credentials for AI. Paid block, see Sponsorship.
- Claude in Google Workspace (Top News, 31,364 likes) — native sidebar via Workspace Marketplace add-on, or connector-based file-open-next-to-chat; Docs (rewrite/restructure without breaking formatting), Sheets (formulas, pivot tables, run Python on a range, write results back), Slides (build from existing layouts, flag issues); big edits surface as accept/dismiss suggestion cards; existing Google sharing permissions carry over. Beta, all paid Claude plans. See Mapping.
- Presented by Sentry — live Oct 14 workshop on debugging agents in production (tracing multi-step agent runs, using user-frustration signals to find broken runs at scale). Paid block, see Sponsorship.
- OpenAI math manuscripts (Top News, 18,378 likes) — 722 manuscripts from an unreleased internal model, ~4,000 attempted problems, 372 Lean-verified paper families; includes progress on the irrationality exponent of pi, a Hodge Conjecture result for a specific object class, and Riemann zeta zero-free-region results. Released after 25 Fields Medal winners signed an open letter on unsupervised AI math races; OpenAI consulted an independent mathematician advisory group. Apache-2.0, public repo. Not deep-fetched — no RDCO hook.
- Signals (quick list, not deep-fetched): (1) Anthropic opens a Claude Startups program, $45,000 in credits and tools; (2) Precisely AI Studio — ready-made AI apps/agents/skills over AI-ready data (sponsor block, see Sponsorship); (3) Matt DesLauriers ships a faster open-source browser vector-graphics renderer; (4) a new 8B continuous-diffusion model matches autoregressive LLMs on math/code; (5) an uncensored GLM-5.3-Flash fine-tune with safety filters stripped, for red-teaming; (6) a new open-source tool to monitor Claude Code and Codex agents in real time — adjacent to RDCO's own always-on agent-monitoring posture but not deep-fetched this issue.
⚠️ Sponsorship
Three paid "Presented by" blocks this issue. Teleport (MCP auth/production report) and Sentry (agent-debugging live workshop) are both already-confirmed members of AlphaSignal's rotating sponsor pool from prior issues. Precisely (AI Studio apps/agents/skills pitch, Signals #2) is a new pool entrant not previously logged — confirmed rotating pool now 31+ distinct sponsors. The masthead "In Partnership with ·" slot is present but unresolved in plaintext (no legible entity name before "Today's Author"), consistent with the majority-unresolved pattern already logged across prior issues. Not counted as a sponsor here since it can't be identified. All three disclosed placements are standard paid ads with no deeper financial tie to AlphaSignal disclosed beyond the placement.
Related
- [[2026-10-06-alphasignal-reflection-beam-watermark-vals-materials-agents]] — immediately preceding issue's open-weight MoE entry (Reflection Beam), same trend line as Mistral Large 4 here
- [[2026-09-17-alphasignal-claude-docs-slides-design]] — direct precedent for Claude's native Google Workspace / Docs-Slides-Design integration push, now extended to a Workspace Marketplace sidebar
- [[feedback_mcp_install_security_review_default]]
- [[project_credibility_for_phdata_sales]]
Checked for overlap: no other same-day newsletter note in the vault shares named items with this issue.