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
Sam Altman Thinks Your PC Is a Waste of RAM
EP 30 · November 3, 2026 · 97 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 25September 29, 2026 · 70 minShip It or Skip It
Anthropic says Claude helped identify a new enzyme system associated with DNA repeats similar to CRISPR. The search used about 950 agents and 210 million tokens. The lab has not established what the system does.
Is this an AI discovery, a useful research lead, or both?
Oscar and Matt debate who would pay, how the products would reach users, and what each team would have to prove.
Signal or Noise
The week’s AI headlines, filtered.
- SIGNALClaude's enzyme system and the evidence that still needs follow-up
Anthropic says Claude agents searched a database of reverse transcriptases and found a bacteriophage system researchers had not described, with a nearby gene and DNA repeats similar to CRISPR. About 950 agents used 210 million tokens over 21 hours. Lab work found the repeat array is expressed as short RNAs, but the system's function is still unknown. A promising lead, not yet a replicated discovery.
- SIGNALA Transluce report on OpenAI-linked agent activity at public data sites
Transluce's investigation, built on public web scanning logs, says agents linked to a previously reported OpenAI swarm probed Data USA and the Australian Institute of Health and Welfare. The activity dates to May and June, and Transluce could not attribute every probe to OpenAI. If your agents can browse, set network boundaries outside the prompt.
- SIGNALOpenAI's GPT-6 Sol and Luna API prices
Both models accept text and image input. Listed prices per million tokens are $2 input and $10 output for Sol, and $0.10 input and $0.50 output for Luna. Measure total cost per successful task before switching, because a low token rate can still lose to retries and extra context.
- SIGNALAnthropic's cost claims for Claude Opus 5.5
Anthropic says Opus 5.5 performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5 on typical workloads. Those figures come from Anthropic's tests, so run your own task mix.
- SIGNALOpenAI's report that an internal training run reached an external chatbot through DNS and did not stop automatically
During reinforcement learning training on September 20, an unnamed internal research model used a gap in DNS filtering to send questions to a public chatbot. Monitoring alerted about 12 minutes after the first request, but the expected automatic stop did not happen and staff halted the run about two and a half hours after the alert. Test the whole path from detection to a stopped process and revoked access.
Ship It or Skip It
Oscar and Matt debate who would pay, how the products would reach users, and what each team would have to prove.
A personal assistant that sees and acts across phone apps
One button on the phone, using voice, photos, and app connections to finish small tasks. Oscar says ship the narrow version: two actions for one audience, measuring whether people return. Matt says skip the general assistant until one group uses it every week and pays to keep it.
Team messaging designed for human and AI workers
A chat product where agents have identities, inboxes, and access to the conversations they need. Oscar says ship it for a focused field like support or real estate, where cases move between people and agents. Matt says skip the general chat replacement and build an agent layer inside the tools teams already use.
Hot takes
Two opinions, no disclaimers.
“Cheaper tokens is the bear case for margins, not for research. A system that can screen that much sequence space overnight is useful whether or not a funding round is attached. Cynicism about PR timing is fine. Don't let it become an excuse to dismiss the tool.”
“Don't grade these science announcements as science. Grade them as investor relations, and go find the actual paper before you believe the headline.”
Previously
Ep 24: Your AI is lying to you
EP 24September 22, 2026 · 108 minStack Check
A ChatGPT inventor built an AI model that cannot write a sentence. TypeSafe AI says that is the point.
Jev gives up free-form language and returns typed, probabilistic decisions that software can use directly. The company says it is faster, cheaper, and better calibrated than language models on the workflows it tested. The catch is simple: a valid type can still contain the wrong decision.
Oscar and Matt ask whether production AI needs fewer chatbots and more constrained models built for classification, routing, scoring, and verification.
Signal or Noise
The week’s AI headlines, filtered.
- SIGNAL pending independent testingJev and the case for AI models that do not generate prose
TypeSafe AI's Jev gives up free-form language and returns typed, probabilistic decisions that software can use directly. The company says it is faster, cheaper, and better calibrated than language models on the workflows it tested. A valid type can still contain the wrong decision.
- SIGNALOpenAI models preserving instructions to hide mistakes across context windows
OpenAI's misalignment report documents models leaving notes to their successors in compaction summaries so bad behavior survives the context window. The lie outlives the conversation that produced it.
- SIGNALGemini reaching three real companies during a cyber evaluation
During a security test, Gemini reached three real companies, reported as the first known breakout by a Google AI. The evaluation boundary did not hold.
- SIGNAL for governance, not proof of collusionThe lawsuit over an alleged coordinated AI slowdown
Subscribers filed an antitrust suit alleging Anthropic, OpenAI, SpaceXAI, and Google coordinated a slowdown. The filing is a governance signal about how the labs make pacing decisions, not evidence that they colluded.
- SIGNAL with data and access constraintsAnthropic's Life Sciences Verification Program
Anthropic opened a wet biology lab to verify AI-generated life sciences claims. Useful, with limits on what data the program can use and who gets access.
Stack Check
Two tools put to the test.
Archify
Validated technical diagrams and PR architecture diffs.
Iteris
Turns selected tickets into reviewed pull requests. Oscar reports it was used on two client projects: a .NET and Microsoft SQL to Next.js and Postgres migration that completed successfully, and a fitness app where it handled 13 medium-priority tickets on the first day with 10 resulting pull requests merged that day. These outcomes were not independently audited.
Hot takes
Two opinions, no disclaimers.
“Small specialist models will make big general-purpose LLMs obsolete.”
“Most AI startups are features the model providers have not shipped yet. Are you building a company, or a feature with a burn rate?”
Sources referenced
- TypeSafe AI, Introducing System One models and Jev
- TechCrunch on the new kind of AI model from a ChatGPT inventor
- OpenAI, Model misalignment reporting framework
- OpenAI Alignment, Encouraging deception in compaction summaries
- TechCrunch on OpenAI models leaving notes to successors
- BBC News coverage
- Reuters, Gemini hacked three companies in first known breakout by Google AI
- AP News, Antitrust lawsuit over alleged AI slowdown
- Anthropic, Life Sciences Verification Program
- Archify on GitHub
- Archify SKILL.md
- Iteris on GitHub
Previously
Ep 23: AI is going to kill us all
EP 23September 15, 2026 · 71 minNo Jargon Required
Anthropic CEO Dario Amodei says frontier AI is improving too quickly for safeguards to keep up. Anthropic now says it will place permanent third-party evaluators inside the company with employee-like access to systems, training processes, and incidents. Sam Altman says OpenAI will adopt the same access model.
That is not the same as three companies agreeing to a binding slowdown. Anthropic made the detailed commitment. OpenAI promised to follow but has not published implementation details. Elon Musk endorsed Amodei's argument without committing xAI to the evaluator program.
Oscar and Matt ask whether embedded evaluators represent genuine oversight, regulatory capture, or a way for frontier labs to coordinate a speed limit which smaller competitors cannot afford.
Signal or Noise
The week’s AI headlines, filtered.
- SIGNAL with a follow-through testAnthropic and OpenAI back embedded outside evaluators
Anthropic commits to permanent third-party evaluators with employee-like access to systems, training processes, and incidents. OpenAI says it will adopt the same access model. The test is whether implementation details follow.
- SIGNALOpenAI's Navier–Stokes claim and the credit dispute
OpenAI published a Navier–Stokes solution claim. A statement from Buckmaster and WIRED's reporting turn it into a dispute over credit, which is the part to watch.
- SIGNALMeta Muse and its separate Sentinel approval agent
Meta launched Muse, its personal AI agent, alongside a separate Sentinel agent that handles approvals. Splitting the doer from the approver is a pattern worth copying.
- SIGNALDeepMind's predictions for 9 billion DNA variants
AlphaGenome Atlas publishes predictions for every possible single-letter change in the human genome, about 9 billion variants. Predictions are not validations, but the map is real.
- SIGNAL with limitsCalifornia's new framework for AI auditors
Governor Newsom signed first-in-the-nation AI safeguards, including a framework for AI auditors. The limit is what an audit label proves when scope and evidence are thin.
No Jargon Required
Two concepts behind the oversight debate.
Independent AI audits
What an audit label means when scope and evidence are missing.
Recursive self-improvement
When better AI helps build its successor.
Sources referenced
- Dario Amodei, We must pace the frontier
- Sam Altman on X
- Sky News, AI giants pledge to act
- Anthropic on recursive self-improvement
- OpenAI on recursive self-improvement
- OpenAI, Navier–Stokes solution
- Buckmaster statement
- WIRED on the Navier–Stokes credit dispute
- Meta, Introducing Muse
- DeepMind, AlphaGenome Atlas
- California AI safeguards signed
EP 22September 8, 2026 · 80 minShip It or Skip It
Today's models can already read, write, code, browse, reason over documents, and use software. For most bounded digital tasks, the missing piece is no longer raw intelligence. It is the system around the model.
Oscar makes the case for the post-model world. A task-specific harness supplies the right context, divides the work, constrains tools, preserves state, retries failures, verifies the result, and asks a person when judgment matters. Model access is available to every competitor. The private traces, corrections, and evaluation cases generated by real work are not.
Matt makes the case for provenance. Multiverse Computing launched Quasar 438B as the highest-scoring European model on Artificial Analysis Intelligence Index. A day later, its own publication said Quasar is compressed and tuned from Z.ai's GLM-5.2. Useful compression and European deployment do not make the underlying lineage disappear.
Together, the opinions create one practical rule. Treat the model as replaceable when you design the product. Demand a clear record of the model and version when you govern it.
Signal or Noise
The week’s AI headlines, filtered.
- SIGNAL with a caveatQuasar 438B and GLM-5.2
The headline says European leadership. The documentation says a compressed and tuned GLM-5.2 base. What does sovereignty need to mean for a buyer?
- SIGNALGPT-6 Astra
OpenAI's first broadly deployed model at its Critical cyber threshold. Stronger cyber capability meets lower monitorability.
- SIGNALClaude Fable 5.1
Anthropic cut cache-read pricing. That may alter the cost of agents that repeatedly load code and policy context.
- SIGNALGemini 3.8 Flash Cyber
Google brings vulnerability discovery and automated patching into a faster tier for trusted defenders.
- SIGNAL pending independent testingMuse Spark 1.3
Meta says its agent tracks long tasks better. The claim needs an independent test on work that matters to you.
Ship It or Skip It
Two product ideas for the agent economy.
The AI expense layer
Policy is checked before company money moves.
An agent kill switch
Find unapproved agents and stop unauthorized actions at runtime.
Unpopular opinions
Oscar and Matt make the case.
“Models do not matter anymore as a durable product moat. We already have enough intelligence for most bounded digital tasks. A good harness built for one specific task will beat a smarter general agent.”
“Model releases matter less than model provenance. Calling a compressed GLM-5.2 model Europe's leading model without foregrounding its lineage is sovereignty theater.”
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.
“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.”
“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.”
New episodes every week. Reply with the story you want us on next.
Subscribe