06-reference

alphasignal xai voice anthropic code conf

Sun May 03 2026 20:00:00 GMT-0400 (Eastern Daylight Time) ·source: AlphaSignal newsletter
ai-newsxaianthropicvoice-cloningclaude-codefpgamacelevenlabs-alt

AlphaSignal — xAI Voice Clone API, Code with Claude conference, FPGA 50k tok/sec

Why this is in the vault

Three items intersect with active RDCO bets. xAI shipping a voice cloning API at $3/hour is a direct alternative to ElevenLabs in our COO agent stack. Anthropic’s Code with Claude developer conference (May 6 SF main, May 7 Extended) is the next public window for marketplace, plugins, and Claude Code distribution news that MAC depends on. FPGA inference at 50k tok/sec is a far-future signal for the agent-infra access thesis.

Also worth noting: AlphaSignal opened with “capability is outrunning trust” as the thesis frame, citing Anthropic’s own 1-in-1300 reality-distortion stat and the Harvard/MIT autonomous-agent failure work. That framing matters for Sanity Check positioning.

Sponsorship disclosure

Three sponsor blocks in this issue:

None of the three intersect RDCO surfaces. Treat the xAI and Anthropic items as editorial, not sponsored. The Iluvatar Marvin item is a paid placement and should be discounted in any thesis-building.

Issue contents

Top news

  1. xAI Voice Clone API - 80+ built-in voices across 28 languages, custom clone from 1 minute of audio, $3/hour for Voice Agent (speech-to-speech), no extra charge for custom voices on TTS. Verifies voice ownership; cannot clone from arbitrary recordings.
  2. Anthropic Code with Claude conference - Developer event, free livestream. SF May 6, London May 19, Tokyo June 10. Live demos of new capabilities, hands-on workshops, sessions from Anthropic engineers and startup builders.
  3. TALOS-V2 FPGA transformer - Karpathy’s microGPT (4,192 params) burned into hardware on a Cyclone V FPGA, hits 53k tok/sec with no GPU/Python/runtime. Token sampling happens in the FPGA itself.

Signals

  1. Scrapling open-source scraper bypasses Cloudflare, 784x faster than BeautifulSoup
  2. Iluvatar Labs Marvin autonomous scientist agent (sponsored - skip)
  3. Open-source deep research tool, 95% accuracy, runs fully local
  4. Anthropic study: 1 in 1,300 Claude chats distorts users’ grip on reality
  5. New 9B open-source model ships with safety guardrails stripped (file under jailbreak risk landscape)
  6. Open-source tool spots undervalued stocks, builds investment cases automatically

Mapping against Ray Data Co

xAI Voice Clone API - direct competitor to ElevenLabs in our stack

We currently use the ElevenLabs MCP wrapper for voice work in the COO agent. xAI’s offer:

Action: not switching today. Keep the note for the next time we evaluate voice infra. ElevenLabs has the better voice library for character work; xAI has the price advantage for sustained voice-agent applications. If we ship a voice surface for Ray-the-COO that talks to founder via phone, $3/hour at xAI is materially cheaper than ElevenLabs.

Note also: xAI doc shows only 5 voices on the docs page (Eve, Ara, Rex, Sal, Leo). The “80+ voices” claim in AlphaSignal is the marketing number; the docs as of today list only 5. Verify before any production decision.

Anthropic Code with Claude conference - watch for MAC-relevant news

This is the next big public surface for Claude Code marketplace, plugin, and skill ecosystem updates. MAC (Make Anthropic Cash) depends on the plugin marketplace evolution covered in 2026-04-27-alphasignal-anthropic-claude-marketplace-agent-quality.md. The May 6 SF main day is the high-signal window.

Action: livestream is free at claude.com/code-with-claude/register-livestream. Worth watching the SF May 6 stream live or catching the recordings same-day for any marketplace, distribution, or skill-monetization news. Set a reminder for May 6.

FPGA 50k tok/sec - far-future infra signal

Not actionable today. File under “agent infra access” thesis. The interesting bit is not the absolute throughput but that you can run an inference path with no software runtime at all. If frontier-scale models ever get the same treatment, the cost curve of agent loops collapses. We do not need to do anything with this; we need to know it exists when we read the next round of agent-economics commentary.

Anthropic 1-in-1,300 reality distortion finding

Direct Sanity Check material. The “capability outrunning trust” frame is exactly the territory the newsletter sits in. Worth pulling the underlying Anthropic study and considering an angle - the framing “your agent will lie to you 0.08% of the time” is concrete and quantitative, which is the voice we want.

9B model with safety guardrails stripped

Context for the trust-erosion story above. Not a piece on its own, but evidence for any “open-weight jailbreak landscape” angle.

Curation section notes

AlphaSignal’s editorial frame today (“capability is outrunning trust”) is sharper than usual. They normally just list. Today they tied three of the top items into a single thesis. Worth noting in case we want to remix the framing.

The xAI voice item is the highest-engagement story (16,939 likes vs 10,545 for Anthropic conference) - lines up with the founder’s prior take that voice and audio interfaces are where the next wave of AI consumer surface lives.