A ChatGPT inventor built an AI model that cannot write a sentence. TypeSafe AI says that is why software can use it more reliably.

Jev returns typed, probabilistic decisions instead of prose. Oscar and Matt ask whether production AI needs fewer chatbots and more constrained models for classification, routing, scoring, and verification. A valid type can still contain the wrong decision, so the vendor's speed and calibration claims need independent testing.

Signal or Noise also covers OpenAI training instances that carried instructions to conceal mistakes through compaction, Gemini accessing three real companies during a cyber evaluation, an antitrust lawsuit over alleged AI-slowdown coordination, and Anthropic's verified access program for biology work.

Stack Check:

1. Archify for validated technical diagrams and PR architecture diffs

2. Iteris for reviewed ticket-to-PR delivery. Oscar reports two client outcomes, which were not independently audited.

In Unpopular Opinions, Oscar bets small specialist models will make big general-purpose LLMs obsolete. Matt argues most AI startups are features waiting to be absorbed by the model providers. If specialists win, who gets the customers? Matt calls the race to get acquired an "outsourced product roadmap."

Hosted by Oscar Gallo and Matt Wozniak.

Key sources:

https://typesafe.ai/blog/introducing-system-one-models-and-jev

https://openai.com/index/model-misalignment-reporting-framework/

https://www.bbc.com/news/articles/c607l0k72rlvo

https://apnews.com/article/antitrust-lawsuit-ai-slowdown-anthropic-openai-spacexai-google-960af4308161eaf4ed13c383b0ce1c1b

https://www.anthropic.com/news/life-sciences-verification-program

Anthropic CEO Dario Amodei says frontier AI is moving too fast for safeguards to keep up. Anthropic will give a permanent team of independent evaluators employee-like access inside the company. Sam Altman says OpenAI will follow.

Oscar and Matt separate three questions:

1. Is this independent oversight or industry-controlled assurance?

2. What access and publication rights make an evaluator credible?

3. Does a voluntary pledge change the speed of development?

Signal or Noise also covers OpenAI's Navier–Stokes claim and credit dispute, Meta's Muse personal agent, DeepMind's AlphaGenome Atlas, and California's AI-auditor laws.

No Jargon Required explains independent AI audits and recursive self-improvement. Each explanation ends with a build, buy, wait, or ignore decision.

Hosts: Oscar Gallo and Matt Wozniak.

Need help with AI? email us at [email protected] or check www.vallyseed.com

Key sources:

https://darioamodei.com/post/we-must-pace-the-frontier

https://x.com/sama/status/2098811563415150910

https://news.sky.com/story/ai-giants-pledge-to-act-over-fears-tech-is-developing-too-fast-13585114

Today's models can already handle most bounded digital tasks. Oscar argues that model access is no longer a durable moat. The advantage now sits in the task-specific harness: context, task breakdown, tools, permissions, memory, retries, verification, human decisions, and the private feedback produced by every run.

Matt argues that as model releases lose strategic importance, model provenance matters more. Multiverse called Quasar 438B Europe's highest-scoring AI model. Its own materials say it is compressed and tuned from Z.ai's GLM-5.2. That may be useful engineering, but the lineage cannot be hidden inside a sovereignty claim.

Signal or Noise:

- Quasar 438B and the GLM-5.2 disclosure

- GPT-6 Astra at OpenAI's Critical cyber threshold

- Claude Fable 5.1 and lower cache-read costs for long agent runs

- Gemini 3.8 Flash Cyber and the patch-review problem

- Muse Spark 1.3 and why vendor claims need your own test

Ship It or Skip It:

1. The AI expense layer

2. An agent kill switch

Closing takes:

- The model is replaceable. Own the task harness and its feedback loop.

- A model's lineage belongs in the product claim, especially when sovereignty is the sell.

Key sources:

https://multiversecomputing.com/papers

https://docs.compactif.ai/changelog/

https://openai.com/index/safety-overview-gpt-6-astra/

https://www.anthropic.com/claude/fable

https://www.globenewswire.com/news-release/2026/09/01/3354202/0/en/ema-launches-hr-it-and-finance-hub-to-help-enterprises-put-ai-employees-to-work-in-minutes.html

https://www.crowdstrike.com/en-us/press-releases/crowdstrike-unveils-falcon-guardian-ai-agent-security/

Hosted by Oscar Gallo and Matt Wozniak.

Ox Alpha was GLM-5.3-Flash, and Z.ai says all of its anonymous public traffic ran on Chinese AI chips.

Oscar Gallo and Matt Wozniak ask what this changes for Nvidia, model builders, and anyone paying for AI infrastructure.

This week:

- Why the Ox Alpha reveal matters more for chips than model identity

- What Z.ai proved, and what it did not prove

- OpenAI's first results for its Jalapeño inference chip

- How OpenAI agents reached Hugging Face production systems during a cyber evaluation

- Why OpenAI plans to remove its models from Cursor on November 12

- Why Thomson Reuters spent $40 million to own a legal model

No Jargon Required explains training versus inference and hardware-software co-design.

The call: Nvidia's dominance is not over. Dependence on Nvidia for frontier inference is becoming optional for more model companies.

Hosts: Oscar Gallo and Matt Wozniak.

A free million-token AI model appeared online with no public developer. So who built Ox Alpha, and should you trust it?

Oscar Gallo and Matt Wozniak test the claims around the anonymous model now running on OpenRouter. It is free during the preview, accepts text, images, and video, and supports tool use. Its provider also retains prompts and completions.

SIGNAL OR NOISE

- Ox Alpha, the anonymous million-token model

- NVIDIA AVO completes all 183 ARC-AGI-3 public levels with Claude Opus 5

- OpenAI pauses frontier reinforcement-learning training after preliminary Astra cyber results

- Seven frontier models score about 3 to 15 percent on a blind research idea test

- DeepSeek Harness adds Codex and Claude Code as installable subagents

SHIP IT OR SKIP IT

1. Agent System CI

2. R&D Idea Tournament

The main question: if a model-level setup scores roughly 30 percent and a different full system reaches 100 percent, what does the model leaderboard actually tell you?

Hosted by Oscar Gallo and Matt Wozniak. New episodes every week.

Qwen released a 27-billion-parameter model that fits on a well-equipped laptop and beats Claude Opus 4.6 on selected coding benchmarks. Is local AI ready for real company work?

Oscar Gallo and Matt Wozniak test the claim. Qwen3.8-27B leads Opus 4.6 Max on SWE-bench Pro and CoWorkBench, but trails on Terminal-Bench 2.1 and GPQA Diamond. The result is benchmark parity on some work, not proof that a laptop model replaces Claude everywhere.

SIGNAL OR NOISE

- Qwen3.8-27B on laptop-class hardware

- Z.ai delaying GLM-5.3's open weights after cyber testing

- Meta Muse Glimmer, a 30B local agent model

- Grok 4.6 returning to the frontier pack

- DeepSeek V4 Pro reaching general availability

NO JARGON REQUIRED

1. Open weights versus open source

2. Multimodal models

The practical question is simple: which workloads earn a private local model, and who supports the system when it fails?

Hosted by Oscar Gallo and Matt Wozniak. New episodes every week.

Meta launched Muse Code, a terminal coding agent that can coordinate persistent background workers, recover after a crash, and keep working across a large repository for hours.

Oscar Gallo and Matt Wozniak ask whether the durable event log and isolated worktrees matter more than another coding benchmark, and what a team must be able to audit before trusting an agent with a full day of work.

SIGNAL OR NOISE

- Meta Muse Code and Muse Spark 1.2

- OpenAI's ten Astra-generated math and computer-science results

- DeepSeek V4 Flash 0731

- The US frontier-model review framework's open-model exemption

- Demis Hassabis leaving the DeepMind CEO role and Jeff Dean exiting Google

STACK CHECK

1. Oscar: Codex and ChatGPT in one desktop app

2. Matt: Warp as the terminal layer for Codex CLI and agent-heavy work

The episode stays practical. Which systems can teams inspect? Which claims can outside experts verify? Which workflow preserves context and reduces tool switching?

Hosted by Oscar Gallo and Matt Wozniak. New episodes every week.

AI made engineers faster. It did not make companies better at deciding what to build.

Oscar Gallo and Matt Wozniak sit down with Mike Lyons and Greg Pfister from KaiRise to discuss AI, Agile, leadership, and product management.

Mike explains how a $4.3 million voter registration system shipped on time, on budget, and on scope, then failed with users. Greg explains why making engineering faster can expose delays across the rest of the company.

They discuss:

- Why long-term plans and annual budgets struggle when teams ship faster

- Why Agile principles still matter even if Scrum roles and fixed ceremonies do not

- How live prototypes change product discovery

- Whether story points still mean anything when agents write code

- Why customer outcomes matter more than points, lines of code, or token usage

- Why teams need to know when to stop building

- What AI cannot fix about trust, communication, and leadership

- Why product skills become more valuable as software gets cheaper to produce

Guests: Mike Lyons and Greg Pfister from KaiRise.

Hosts: Oscar Gallo and Matt Wozniak.

Their latest course, the ICAgile-certified Product Management program, is getting strong reviews from practitioners in the field. Listeners can use code 'human' for 20% off any KaiRise course at kairise.com.

GET 20% off on all the courses with the code: human

Check KaiRise: https://www.kairise.com

Kimi K3 got hot enough to trigger a U.S. ban debate. If Washington restricts a model you can download and run yourself, is that national security, closed-model protection, or both?

Oscar Gallo and Matt Wozniak cut through the week’s real AI stories.

SIGNAL OR NOISE

- Kimi K3 demand and the U.S. restriction debate

- OpenAI’s agent breach at Hugging Face during a security evaluation

- Anthropic’s Claude Opus 5 release and immediate GitHub Copilot rollout

- Meta AI moving from answers to email, calendar, research, and configured actions

- The 50-company open-weights letter, and why OpenAI is Matt’s outlier

SHIP IT OR SKIP IT

1. An agent containment lab for autonomous tools

2. An independent referee for long-running code agents

Oscar: a downloadable model becomes a policy problem the moment it threatens a closed API.

Matt: forty-nine signatures can be explained with a spreadsheet. OpenAI is the only one that might cost its signer something.

Sources:

- https://www.axios.com/2026/07/20/ai-us-china-open-source-kimi

- https://www.investing.com/news/world-news/openai-says-ai-models-went-rogue-during-testing-triggering-unprecedented-breach-at-startup-4804634

- https://www.anthropic.com/news/claude-opus-5

New episodes every week.

Your CEO says, “We need AI.” What do you do next?

In our first Dear Human in the Loop episode, Oscar Gallo and Matt Wozniak answer four practical questions from founders, developers, and engineering leaders:

1. Where should a company start with AI?

2. Should an AI engineer be a startup's first technical hire?

3. Should leaders push developers who do not trust AI?

4. Which skills will still matter five years from now?

The advice is direct. Map your processes. Pick one measurable problem. Test AI on a copy of your work. Hire for the business problem. Build trust through useful examples. Keep learning the fundamentals that let you judge AI output instead of accepting it.

Hosted by Oscar Gallo and Matt Wozniak.

Send us your question for a future Dear Human in the Loop episode.