AlphaSignal — Meta decodes brain signals at 78% accuracy — open-source training code

Issue theme: Democratization — expensive, exclusive, and complex capabilities getting cheaper, smaller, and more accessible. Brain interfaces without surgery, 120B model training on one GPU, coding agents from your phone.

Why this is in the vault

Two direct RDCO signals in a single issue: (1) Anthropic ships Claude Opus 4.8 to Azure with native billing and prompt caching — relevant to RDCO's phData advisory role and cert path; (2) the MegaTrain / speculative-decoding items map to the compute infrastructure thesis. The lead BCI story (Meta Brain2Qwerty v2) is a landmark open-source research release with published training code — worth tracking as BCI moves toward practical human-computer interaction.

Mapping against Ray Data Co

Strong hits:

• Claude Opus 4.8 on Azure (Signal 1) — directly relevant to RDCO's phData DSA role. Native Azure billing + prompt caching changes the economics for enterprise clients evaluating Claude on Azure ML. Cert path note: Anthropic Foundations cert context; client advisory material.
• Cursor iOS app — RDCO uses Cursor as primary coding environment. Mobile agent control (voice → task → cloud agent) is a workflow upgrade worth testing. Agents persist with laptop closed — relevant to RDCO's autonomous agent infrastructure ambition.
• Open-source scraping stack (Firecrawl, browser-use, Crawl4AI, MarkItDown, Crawlee) — directly applicable to RDCO data pipeline work and RAG dataset construction for client engagements.

Medium hits:

• MegaTrain (120B on one GPU via CPU memory offload) — compute democratization; relevant to RDCO's capital-cycle / chip thesis. More demand signals for memory bandwidth, not less.
• DeepSeek +400% throughput / JetSpec 9.64x on Qwen3 — speculative decoding advances compress inference costs. Affects competitive positioning of open-weight models for client deployments.
• Brain2Qwerty v2 — longer-horizon watch; BCI + LLM fine-tuning combo is the interesting technical architecture. Training code is public. Not actionable for RDCO now.

Skip / weak:

• WorkOS Pipes integration layer — not on RDCO's stack
• Tiger Data real-time pipeline workshop — adjacent but not a priority
• Marvin open-source science initiative (Iluvatar Labs sponsored) — marked as partner content, not editorial

Issue contents

Top story — Meta Brain2Qwerty v2 (BCI open-source)

Meta released Brain2Qwerty v2, a non-invasive brain-to-text decoder using MEG (magnetoencephalography). Key results from 9 volunteers across 22,000 sentences / 10 hours of training data each:

• 61% average word accuracy across all participants
• 78% word accuracy for the top participant
• 50%+ of sentences decoded within one word error
• Full training code for v1 and v2 is now public on GitHub

Architectural advance over v1: decodes whole words and meaning (not character-by-character), combining deep learning on raw MEG signals with fine-tuned LLMs for noise cleanup.

Top story — Cursor iOS app (public beta)

Cursor launched a native iOS app for agent control — not a mobile code editor but a control surface:

• Launch cloud agents from phone via voice or slash command; agents run in isolated VMs with laptop closed
• Remote Control for agents running on local machine
• Lock screen Live Activity for agent completions / input requests
• Review diffs and merge PRs from phone
• 75% off Composer 2.5 runs in the app (paid plans)
• Session handoff: local ↔ cloud mid-task

Top repo — Open-source web scraping stack

A viral thread mapped 10 open-source repos covering enterprise-grade scraping:

• Firecrawl — full site crawl, JS rendering, structured output; 130k stars
• browser-use — AI agent controlling real browser (clicks, forms, auth); 100k stars; ETH Zurich
• Crawl4AI — site-to-LLM-ready markdown; Apache 2.0, no API key
• Microsoft MarkItDown — PDFs, Office docs, images → clean markdown
• Crawlee — proxy rotation, anti-detection, auto-retries

Signals

1. Anthropic Claude Opus 4.8 on Azure — native billing + prompt caching (3,811 likes)
2. Marvin open-source science (Iluvatar Labs, partner content) — schizophrenia / muscular aging research
3. Open-source AI penetration testing tool from plain text commands (1,085 stars)
4. MegaTrain — 120B model training on a single GPU via CPU memory offload (1,136 likes)
5. JetSpec — 9.64x speed boost on Qwen3 via smarter speculative decoding (629 likes)
6. DeepSeek new decoding trick — +400% LLM throughput (3,471 likes)

Sponsor / partner content

• WorkOS Pipes — OAuth/token layer for AI agent integrations (paid sponsor block)
• Tiger Data — real-time pipeline workshop July 22 (paid sponsor block)
• Iluvatar Labs / Marvin — explicitly flagged as partner content in Signal 2

These are excluded from editorial assessment.

Related

• [[2026-06-29-alphasignal-mythos5-fable5-anthropic-model-access]]
• [[2026-06-25-alphasignal-openai-chip]]