"Thinking Machines Inkling 🎙️, OpenAI GPT-Red Security 🛡️, PrismML 27B" — @AlphaSignal
Why this is in the vault
Three top-news items each hit a different active RDCO bet: open-weight compute demand (Inkling 975B MoE) as a Phase 2 chip-cycle demand amplifier; automated red-teaming's 84% prompt-injection attack success rate as a direct input to RDCO agent security posture; and PrismML Bonsai 27B's on-device story as an efficiency frontier signal. The Ant Ling 1T-Zero signal adds a fourth angle on zero-annotation reasoning capability compression. Dense issue — all four editorial top items are vault-relevant.
⚠️ Sponsorship
Three sponsor placements in this issue:
- Cirrascale Cloud Services — "NVIDIA Blackwell GPUs on a Purpose-Built AI Cloud." Placed between Inkling and GPT-Red items. GPU cloud provider; reports 30% higher job completion rates vs. traditional hyperscalers, flat pricing.
- Tiger Data (TimescaleDB) — "AI Data Piles Up. Postgres Can Keep Up." Placed between GPT-Red and PrismML items. AI data/Postgres analytics vendor; hypertables, 95% storage compression, continuous aggregates.
- HydraDB — "Struggling with context management for your AI agents? Try HydraDB for free." Placed in Signals section at item #2. Graph-native agent context management, sub-200ms recall, single API.
Editorial items appear clean — no sponsor-placed curation detected.
Curation section
Thinking Machines releases Inkling — open-weight multimodal (text, audio, image)
- Summary: First open-weights release from Thinking Machines Lab (Mira Murati's company). 975B total parameters, 41B active per task via MoE architecture. 1M-token context window. Trained on 45T tokens across text, image, audio, and video. Native audio and vision with no glue components. Top open-weights audio scores on VoiceBench, MMAU, and AudioMC. Controllable reasoning effort (dial up/down). Apache 2.0, available on Hugging Face. Inkling-Small previewed for lighter workloads.
- Relevance: RDCO relevant — Phase 2 chip-cycle demand signal; open-ecosystem compute acceleration.
- Sponsor/self-promo: None.
OpenAI GPT-Red automated red-teaming cuts prompt injection failures 6x
- Summary: GPT-Red uses self-play to attack OpenAI's own models. Achieves 84% success across attack scenarios vs. 13% for human red-teamers. GPT-5.6 Sol now has 6x fewer failures on the hardest prompt injection benchmark. Independently discovered a "fake chain of thought" attack class before human researchers. Demonstrated against an autonomous vending machine agent (price manipulation, order cancellation). Defenses propagate automatically to all models built on OpenAI's stack.
- Relevance: RDCO relevant — directly informs agent security posture; prompt injection is the #1 agent threat vector; quantifies the attack surface RDCO's Brigade House adversarial review is designed to catch.
- Sponsor/self-promo: None.
PrismML ships Bonsai 27B — open-source model at 3.9 GB, phone-ready
- Summary: Standard 27B requires ~54GB at full precision, ~18GB at 4-bit. PrismML Bonsai ships two variants: Ternary (5.9GB, laptop-ready, 95% of full-precision across 15 benchmarks including math/coding/tool use) and 1-bit (3.9GB, phone-ready, 90% of full-precision, ~10.8 tok/s on iPhone at 672 tokens per 1% battery). Supports image inputs on-device. Enables fully local multi-step AI agents. Apache 2.0, llama.cpp and MLX support.
- Relevance: RDCO relevant — model efficiency frontier signal; on-device agent deployment economics.
- Sponsor/self-promo: None.
Signals
| # | Headline | Metric | Relevance |
|---|---|---|---|
| 1 | OpenAI Codex completes a 9-hour coding task after hitting the usage limit | 3,403 likes | Mildly relevant — agentic long-horizon task capability |
| 2 | [SPONSOR — HydraDB] | — | Sponsor |
| 3 | OpenAI releases CLIP — images matched to text with zero labeled training data | 33,995 stars | Not relevant — prior art, historical model |
| 4 | Ant Ling trains Ring-2.5-1T-Zero: 1T param reasoner with zero human annotations | 492 likes | RDCO relevant — zero-annotation reasoning compression; open ecosystem closing gap on frontier |
| 5 | Nvidia releases quantized 1B embedding model, 34 languages, for RAG | 2,689 downloads | Mildly relevant — RAG infrastructure efficiency |
| 6 | HyperFrames open-source tool generates videos from HTML, CSS, and JS | 1,983 likes | Flag: HyperFrames skills exist in the RDCO Claude Code harness — if this is the same project, this is organic traction coverage; verify |
Mapping against Ray Data Co
Inkling 975B MoE is the clearest Phase 2 chip-cycle demand signal in this issue: a 975B-parameter open-weight model released under Apache 2.0 from Mira Murati's Thinking Machines Lab, a team with the credibility to drive enterprise adoption. Open-weight releases at this parameter scale sustain Phase 2 — every organization that fine-tunes or serves Inkling is buying or renting Blackwell-class GPUs. This broadens the demand side of the chip capital cycle beyond hyperscaler labs, which is exactly what Phase 2 needs to deepen rather than peak.
GPT-Red's 84% attack success rate is the most operationally specific security finding for RDCO agent deployments. The self-play approach independently discovered a "fake chain of thought" attack class and demonstrated price manipulation against an autonomous vending machine agent — the exact threat model the Brigade House adversarial review SOP (08-tooling/2026-07-10-house-adversarial-review.md) was built to catch. The 6x improvement in OpenAI-stack defenses is useful as an external benchmark; RDCO's non-OpenAI-stack agents won't inherit those defenses automatically, which means the 84% attack figure is the baseline to stress-test against.
PrismML Bonsai 27B at 3.9GB closes the on-device loop opened by yesterday's AlphaSignal on-device RL issue. A 27B-class model at phone size makes fully local multi-step agent workflows viable — the privacy and latency economics shift materially when capable reasoning can run offline without a cloud call.
Ant Ling Ring-2.5-1T-Zero signals how fast the zero-annotation training frontier is moving: a trillion-parameter reasoner that achieves competitive performance without human labels. This is directly on the cost-compression curve RDCO tracks; annotation bottlenecks are expensive, and their removal accelerates the open-ecosystem gap closure mentioned in the newsletter lede.
Related
- [[2026-07-14-innermost-loop-singularity-trade-chip-cycle]] — TSMC +68% June revenue confirming Phase 2; contextualizes Inkling's compute implications as demand-deepening rather than demand-peaking
- [[2026-07-15-alphasignal-anthropic-canada-credits-ondevice-rl]] — yesterday's AlphaSignal; on-device RL and model efficiency thread that PrismML Bonsai directly continues
- [[2026-07-07-dram-hbm-phase2-phase3-early-signals]] — Phase 2→3 tripwire monitor; Inkling-class open-weight deployments are Phase 2 demand-sustaining events worth logging against the tripwire checklist