Gemini 3.8 Flash and Flash Cyber land, Fairwind gates the cyber model

AIpositive +0.8#model-release #gemini #safety #google #fairwind

Google ships its third Flash in six weeks: a same-price smarter workhorse plus a defender-only cyber model that patches near frontier quality at Flash cost.

What they showed / shipped

Demis and DeepMind announced Gemini 3.8 Flash and 3.8 Flash Cyber, the third Flash drop in six weeks.
Google DeepMind: 3.8 Flash targets agentic coding and multi-step reasoning; Flash Cyber is for vulnerability detection and automated patching.
Official blog: same intro price as 3.7 Flash at $0.75 / $3.75 per 1M tokens through Dec 31, 2026. Flash Cyber is Fairwind-only for vetted defenders.
Hard numbers from Google: Flash Cyber hits 47.2% pass@1 on CWE-Bench vs 47.8% for a leading frontier model at lower cost, and Chrome Security got 2.6x more correct patches than larger commercial models.
The Verge notes 3.8 Flash works harder and may cost more per task because it takes extra reasoning steps and tool calls.
📌 On the radar: researchers keep pressing OpenAI on Astra safety pacing after yesterday Critical cyber preview. Update only, not a new topic.

Why it matters

Builder lens: a same-price Flash upgrade you can point agents at today, plus a cyber sibling that is actually gated instead of dumpable.
Creator lens: third Flash in six weeks is the cadence story. Pair the price lock with the Chrome 2.6x patch claim.

Sources

Perplexity open-sources Lily, the Mac hybrid-compute engine

Techpositive +0.8#open-source #agents #perplexity #tooling #hardware

Yesterday hybrid compute was a product feature. Today the local inference engine behind it is open source for Qwen on Apple silicon.

What they showed / shipped

Perplexity is open-sourcing Lily, the local inference engine built for hybrid compute in Perplexity Computer, specialized for Qwen3.6-35B-A3B on Apple silicon.
Arav: Lily powers hybrid compute on the Mac app and is out as OSS.
Same day Arav showed Computer running fully local on DGX Spark.

Why it matters

Builder lens: you can study or reuse the exact engine that made cloud-to-local handoff real, not just read the marketing post.
Creator lens: demo sensitive docs staying on-device, then cut to the Lily repo. That is the privacy primitive on camera.

Sources

Meta Muse Spark 1.3 hits the frontier coding pack

AIpositive +0.8#model-release #muse #meta #coding #agents

Meta Superintelligence Labs ships Muse Spark 1.3: Wang says it is competitive with Fable 5.1 and better than GPT-5.6 Sol on coding, with fewer tokens and Contemplating multi-agent mode.

What they showed / shipped

Meta introduced Muse Spark as a natively multimodal reasoning model with tool use, visual chain of thought, and Contemplating multi-agent mode.
Alexandr Wang: Muse Spark 1.3 is competitive with Claude Fable 5.1 and better than GPT-5.6 Sol on coding, same API price as 1.2, about 25% fewer tokens.
Artificial Analysis puts Muse Spark 1.3 (xhigh) at 61 on the Intelligence Index, tying GPT-5.6 Sol and Grok 4.6, with 235 tok/s output speed.
Reddit singularity thread flagged the surpassed Fable 5 / GPT-5.6 Sol claim as it spread.
📌 On the radar: Fable 5.1 is already one-shotting games and 3D worlds, and minchoi says it beats GPT-5.6 Sol Max on CursorBench 3.2 at ~38% lower cost.

Why it matters

Builder lens: Meta finally has a coding/agent model in the same band as Fable and Sol, not a generation behind.
Creator lens: chart the AA score plus the 235 tok/s speed. Speed is the teachable edge vs Fable.

Sources

AI answer engines are citing fake best-of pages and missing numbers

Technegative -0.6#data #hype-cycle #perplexity #skeptic #public-perception

Two audits land the same day: 215k manufactured software pages feeding Perplexity, and about a third of Perplexity number citations that do not contain the number.

What they showed / shipped

HN: three sites mass-produced 215,128 best software pages that Perplexity cites.
Separate audit: about a third of Perplexity citations do not contain the number they are cited for.

Why it matters

Builder lens: if your product recommends software or quotes stats from an answer engine, you need a cite-verify step, not trust.
Creator lens: this is the perfect skeptic segment. Show a fake best-of page next to a citation that fails a number check.

Sources

Runway ships Dev MCP so coding agents can drive its API

Techpositive +0.7#tooling #video-gen #runway #agents #mcp

Runway exposes Model Routers, tasks, and debug hooks through an MCP server your coding agent can call without leaving the IDE.

What they showed / shipped

Runway announced Dev MCP: connect a coding agent to the Runway developer platform to build, debug, and manage Model Routers and tasks in-place.
📌 On the radar: World Labs Atlas demos keep expanding, including sports bullet-time from a few camera views.

Why it matters

Builder lens: video gen stops being a separate console app and becomes a tool call inside Claude Code / Cursor-class agents.
Creator lens: demo an agent spinning up a router and kicking a generation without opening runwayml.com.

Sources

x/@runwayml X@runwayml
♥ 87↻ 13👁 10K

US government sides with OpenAI in the NYT training fight

Policymixed +0.1#policy #regulation #openai #copyright #public-perception

The Trump administration backs OpenAI against the New York Times: training on copyrighted material is not infringement, which would reshape every lab that scrapes the open web.

What they showed / shipped

TechCrunch: the US government sides with OpenAI on training LLMs on copyrighted material.
The Verge: the administration is supporting OpenAI in the NYT copyright lawsuit.

Why it matters

Builder lens: if courts follow DOJ, the training-data legal overhang shrinks for US labs and anyone fine-tuning on web corpora.
Creator lens: this is the policy beat creators care about. Training is not infringement is the line to put on screen.

Sources

Six curl CVEs after OpenAI and Anthropic found zero

Policymixed -0.2#safety #benchmarks #open-source #skeptic #curl

Aisle re-tested curl after frontier labs reported clean: six new CVEs. AI vuln hunting still misses what a focused human pass catches.

What they showed / shipped

Aisle blog on HN: they found six curl CVEs after OpenAI and Anthropic came back with zero.

Why it matters

Builder lens: do not treat a clean lab scan as a ship gate. Use models for triage, then a human or specialized tool for the second pass.
Creator lens: this is the skeptic counterweight to Flash Cyber hype. Same week Google sells cyber agents, curl still had holes the labs missed.

Sources

aisle/six curl CVEs after labs found zero @goobreee
♥ 152