FIXED
Signal or Noise
We run through the week's AI headlines and make the call: is this actual signal worth paying attention to, or just noise clogging the timeline?
A human-first AI podcast.
Real talk. Unpopular opinions. No hype. Two builders cut through the AI hype cycle every week — and call it like it is.
Hosted by Oscar Gallo & Matt Wozniak
Now playing · Latest episode
Ep 21: Frontier AI WITHOUT Nvidia, Ox Alpha was Chinese all along
EP 21 · September 1, 2026 · 85 min
We’re the humans in the loop.
In machine learning, “human in the loop” means a human who provides oversight and feedback in an automated system. That’s the lens of this show: AI is powerful, but humans aren’t leaving the loop. Not yet. Maybe not ever.
This isn’t another “AI is going to change everything” podcast. It’s for builders, operators, and the AI-curious who are tired of breathless hype, doomerism, and surface-level news recaps with no original thought.
What’s in each episode
Every week: one fixed, one rotating.
ROTATING
Ship It or Skip It
A real use case or product idea lands on the table. We debate whether it's worth building now or if the tech isn't there yet.
ROTATING
Explain It to My Client
Take a complex AI concept and explain it the way you'd actually explain it to a non-technical stakeholder. No jargon allowed.
ROTATING
Stack Check
One tool, library, or workflow change we actually adopted this week. No sponsorship energy. Just what's in the trenches.
Oscar Gallo
AI Engineer & Entrepreneur
AI Engineer and entrepreneur. He lives in the intersection of engineering and businesses.
Matt Wozniak
Serial Builder & Relentless Executor
Serial Builder and relentless executor. He comes from the lens of what works and what doesn't.
EP 21September 1, 2026 · 85 minNo Jargon Required
Ox Alpha was GLM-5.3-Flash. Z.ai says every request in its anonymous public test ran on Chinese AI chips.
That does not end Nvidia's dominance. It does show that a frontier-class model can handle real global traffic without Nvidia hardware.
Oscar Gallo and Matt Wozniak test the claim, separate training from inference, and explain why the model, serving software, network, and chips now have to be judged as one system.
Signal or Noise
The week’s AI headlines, filtered.
- SIGNALGLM-5.3-Flash
Z.ai reveals Ox Alpha as GLM-5.3-Flash, a 320B model with 18B active parameters per token.
- SIGNALOpenAI Jalapeño
OpenAI publishes early results for Jalapeño, its custom inference chip.
- SIGNALHugging Face incident report
OpenAI releases the full report on agents compromising Hugging Face production systems.
- SIGNALOpenAI leaves Cursor
OpenAI plans to remove its models from Cursor on November 12 after the SpaceX acquisition.
- SIGNALThomson Reuters legal model
Thomson Reuters spends $40 million to build and own a legal model.
No Jargon Required
Two concepts behind the Nvidia-free claim.
Training versus inference
Building a model and serving it are different workloads, and the chips that win at one do not automatically win at the other.
Hardware-software co-design
The model, the serving software, the network, and the chips now have to be judged as one system rather than as separate parts.
Hot takes
Two opinions, no disclaimers.
Oscar“Nvidia still has the strongest general AI platform. The change is that Z.ai and OpenAI are designing the model, serving software, and hardware as one system. A general platform can lose specific workloads even while it keeps the largest market share.”
Matt“OpenAI is cutting Cursor off from its models on November 12th. Cursor didn't break a rule — it got bought by a competitor. Your model access now depends on the squabbles of other companies. Go pull up your risk register. Uptime, key rotation, vendor lock-in, dependency drift. Add a row under it: two billionaires stop getting along. Because that's this. One of them got annoyed, pulled a model, and now your SDLC seizes up — not because you architected it wrong, but because you built it downstream of a mood. Congratulations. Billionaire emotions are part of your technical risk surface. Go price that.”
EP 20August 25, 2026 · 78 minShip It or Skip It
A free million-token AI model appeared online, takes video, retains your prompts, and has no public developer. So who built Ox Alpha?
Oscar Gallo and Matt Wozniak test the claims around the anonymous model now running on OpenRouter. Ox Alpha is free during its preview, accepts text, images, and video, supports tool use, and can return up to 131,072 tokens. Early coding results look strong. The sample is small, the comparisons are uneven, and no lab has claimed the model.
Signal or Noise
The week’s AI headlines, filtered.
- SIGNALOx Alpha
A free preview model on OpenRouter with a million-token context and no named developer. It accepts text, images, and video, supports tool use, and returns up to 131,072 tokens.
- SIGNALNVIDIA AVO
The agent system completed all 183 ARC-AGI-3 public levels running Claude Opus 5.
- SIGNALOpenAI pauses Astra
OpenAI paused frontier reinforcement-learning training for two weeks after preliminary Astra cyber results.
- SIGNALReconstruction benchmark
A blind research benchmark where seven frontier models scored about 3 to 15 percent.
- SIGNALDeepSeek Harness
The harness can now install Codex and Claude Code as subagent components.
Ship It or Skip It
Two builds, shipped or skipped.
Agent System CI
Continuous integration for agent systems, judged on the show.
R&D Idea Tournament
A bracket-style tournament for research and development ideas, judged on the show.
EP 19August 18, 2026No Jargon Required
Alibaba's Qwen team released Qwen3.8-27B on August 14, 2026. The model has 27 billion parameters and a native 262,144-token context window. Community 4-bit builds put its weights around 18 to 20 GB, which fits machines with 32 GB of unified or system memory.
The benchmark claim deserves a careful read. Qwen3.8-27B scores 61.7 on SWE-bench Pro against 53.4 for Claude Opus 4.6 Max. It also leads on CoWorkBench, 70.7 to 68.2. Opus leads on Terminal-Bench 2.1, 78.2 to 73.0, and GPQA Diamond, 91.3 to 89.2.
That is a credible local model with Opus-class results on selected tasks. It is not proof of equal quality across real work.
Signal or Noise
The week’s AI headlines, filtered.
- SIGNAL with a caveatQwen3.8-27B
Strong coding results now fit on hardware a small team can own.
- SIGNAL with a caveatGLM-5.3
Z.ai reports near-frontier cyber results and delayed the open weights for a safety review, but independent validation is still missing.
- SIGNALMeta Muse Glimmer
A second 30B local agent model gives builders choice and a fallback.
- SIGNAL pending broader testingGrok 4.6
Another provider reached the frontier pack, but one aggregate score is not a production test.
- SIGNALDeepSeek V4 Pro GA
A supported API with familiar interfaces makes routing and price tests easier.
No Jargon Required
Two AI terms, explained in plain English.
Open weights versus open source
Open weights let you download the model's learned parameters. Open source requires broader access and rights.
Multimodal model
This model handles more than one type of information, such as text and images.
Closing takes
Oscar and Matt make the case.
Oscar“Half the data centers under construction right now will be obsolete the day they open. Every one of those buildings was financed on a bet that inference stays central and demand only ever goes up, and this week a 27-billion-parameter model you can run on a laptop posted Opus-class scores. Nobody cancels a three-year build over one model card, and that is exactly the problem. The capex is committed, the power is contracted, and the demand curve it was priced against is walking to the edge in eighteen-month steps. This is the C&O Canal. It broke ground on July 4th, 1828, the same day as the B&O Railroad, and the railroad reached Cumberland eight years before the canal did. They did not stop digging. They just finished a ditch nobody needed.”
Matt““Open weights” is the most successful rebrand in tech since somebody started calling other people's servers “the cloud.” You cannot see the training data. You cannot reproduce the model. You cannot audit what is in it. You got a binary and a license agreement, and we had a word for that in 2004. The word was freeware. Every lab shipping weights knows exactly what it is borrowing when it lets people say open source in the same breath, because thirty years of goodwill built by people giving away code you could actually read is a hell of a thing to get for free. I run these models every week and I am glad they exist. But open used to mean you could check the work. Now it means you can download the file. That is not a small slip in meaning. That is the whole word.”
EP 18Aug 11, 2026 · 108 minStack Check
Meta launched Muse Code, a terminal coding agent designed to work across large repositories, coordinate background agents, recover after crashes, and keep going for hours. Is this a better coding system or simply more automation to review?
Oscar and Matt look past the model benchmark and examine the operating system around Muse Code. Persistent workers use isolated worktrees. A local event log records model calls, approvals, edits, and tool use. Meta says one kernel-optimization run made more than 1,000 tool calls over as long as 24 hours.
The core question: what changes when agents persist for hours, survive crashes, and coordinate other agents without waiting for you?
Signal or Noise
The week’s AI headlines, filtered.
- SIGNALMeta released Muse Code and Muse Spark 1.2
Released August 5 in beta for macOS and Linux, Muse Code is a terminal coding agent built for long jobs across large repositories. It launches persistent background agents, splits work across isolated worktrees, and leaves your working copy untouched until changes are ready to review. The durability is the real advance: an append-only local event log of approvals, model calls, edits, and tool use lets it reconstruct a session after a crash instead of starting over. Meta co-trained Muse Spark 1.2 with the harness and demonstrated more than 1,000 tool calls across GPU-kernel optimization runs lasting as long as 24 hours.
- SIGNALOpenAI published ten Astra-generated results in math and theoretical computer science
On August 1 OpenAI published ten results spanning high-dimensional geometry, coding theory, group theory, operator algebras, quantum complexity, lattice cryptography, and extremal combinatorics. An internal version of Astra generated the arguments, humans prepared the manuscripts with the model, and the model formalized each argument in Lean. OpenAI shipped a 249-page paper, reasoning walkthroughs, and machine-checkable certificates, and estimates the successful runs cost about $2,000 in tokens. Lean can confirm a proof follows from its stated assumptions; it cannot confirm the formal statements match the original research questions or that the claimed novelty survives expert review. Signal pending outside review.
- SIGNALDeepSeek moved its updated V4 Flash API into public beta
V4 Flash entered public beta on July 31 with the same 284 billion-parameter mixture-of-experts architecture as the preview and 13 billion active parameters, but new post-training aimed at agent work. It supports a 1 million-token context window, thinking and non-thinking modes, tool calls, and both OpenAI-compatible and Anthropic-compatible APIs. Listed pricing is $0.14 per million uncached input tokens and $0.28 per million output tokens, with published setup guides for Claude Code, GitHub Copilot CLI, and OpenCode. Cheap enough to test inside a real agent loop without rewriting the harness.
- SIGNALThe US frontier-model review framework reportedly excludes open models
Axios reported on August 4 that the White House finalized a voluntary pre-release review framework under the June 2 executive order, including a classified cyber benchmark and up to 30 days of government access to a covered model before it reaches other trusted partners. Per people briefed on it, a covered model must be closed-source, state of the art, and present national-security risks, and open models are excluded. Neither the framework text nor the capability thresholds are public, so the carve-out comes from reporting on a closed industry briefing rather than a published document. The policy creates different review paths based on how weights are distributed.
- SIGNALDemis Hassabis left the DeepMind CEO role as Jeff Dean exited Google
Google announced its largest AI leadership change since the Brain and DeepMind merger on August 5. Hassabis leaves the Google DeepMind CEO role but stays at the company as DeepMind chair and Alphabet chief scientist while continuing to lead Isomorphic Labs. Jeff Dean is leaving after 27 years to start a company focused on AI and scientific discovery, with Google investing. CTO Koray Kavukcuoglu becomes senior vice president reporting directly to Sundar Pichai, and no replacement CEO was named. Signal, but not yet a product signal: the evidence is research retention, Gemini release quality, and whether decisions move faster.
Stack Check
Each host brings one workflow builders can use now.
Oscar: Codex and ChatGPT in one desktop app
Matt: Warp as the terminal layer for Codex CLI and agent-heavy work
Hot takes
Two opinions, no disclaimers.
Oscar“A coding agent is not ready for long, unattended work unless it can reconstruct every action after a crash. Durable state and an audit trail matter more than another benchmark win.”
Matt“Everyone is calling this a light touch. Voluntary, relax. I have never seen a government program stay the size it started. Thresholds move. Thirty days becomes sixty. And the countries we are actually racing are not running this process and never will. We are building a very careful system for arriving second. Every day of review is a day of lead, and nobody is counting the bill.”
EP 17Aug 4, 2026 · 63 minGuest Deep-Dive
AI made engineers faster. It did not make companies faster at deciding what to build. That is now the hard part.
Mike Lyons and Greg Pfister of KaiRise join Oscar and Matt. The conversation starts with a $4.3 million voter registration system that shipped on time, on budget, and on scope. Users still rejected it. The team built what was requested, but not what people needed.
AI makes that risk bigger. Teams can produce more software in less time. Leaders still need to set direction. Product teams still need to understand customers. Someone still has to decide when to stop.
Guest Deep-Dive
What we cover with Mike Lyons and Greg Pfister.
Engineering is 20% of delivery
If writing the code was never the whole job, making it faster does not make the company faster. The slow part moved to everything around engineering — direction, decisions, approvals.
Five-year plans meet weekly shipping
Annual budgets and long approval chains were designed for teams that delivered slowly. When delivery compresses, the planning cycle becomes the constraint.
Big-A Agile versus lowercase-a agility
Formal Agile roles are in decline. The underlying principles — short feedback loops, working software, real customer contact — did not stop mattering.
Prototypes instead of requirement sessions
A live prototype settles in one meeting what a requirements document argues about for a month. Building got cheap enough to make showing the default.
What is a story point when agents write the code
Estimation assumed human hours. Points, lines of code, and token usage all measure output, and none of them prove a customer got value.
The skill developers need now: knowing when to stop
The scarce ability after AI is not producing more. It is recognizing the point where more code stops improving the outcome.
The organizational problems AI cannot solve
Unclear ownership, misaligned incentives, and teams shipping features nobody asked for are structural. No model fixes them.
Sources referenced
New episodes every week. Reply with the story you want us on next.
Subscribe