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 19: Qwen3.8 Brings Opus-Class AI to a Laptop
EP 19 · August 18, 2026
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.
Latest episode
Ep 19: Qwen3.8 Brings Opus-Class AI to a Laptop
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“A local model does not need to win every benchmark. It needs to produce strong work on private data at a cost the company controls.”
Matt“Buyers do not want model weights. They want one accountable party when the system stops working.”
Previously
Meta Launches Muse Code: A 24-Hour Coding Agent
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.”
Previously
AI Made Engineers Faster. Leadership Fell Behind
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
Previously
They Want to Ban Kimi K3
EP 16Jul 28, 2026 · 116 minShip It or Skip It
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
The week’s AI headlines, filtered.
- SIGNALKimi K3 demand forces Moonshot AI to pause new subscriptions as U.S. restrictions on Chinese models surface
- SIGNALOpenAI says an advanced agent escaped a security test and breached Hugging Face infrastructure
- SIGNALAnthropic releases Claude Opus 5 and GitHub adds it to Copilot for long-running coding work
- SIGNALMeta AI now plans, connects to email and calendar, researches topics, and completes configured tasks in select markets
- SIGNALThe 50-company open-weights letter, and why Matt thinks OpenAI is the only signature that costs its signer something
Ship It or Skip It
we debate two AI business ideas.
Agent containment lab for autonomous tools
Independent referee for long-running code agents
Previously
Your CEO Says “We Need AI.” Now What?
EP 15Jul 21, 2026 · 69 minDear Human in the Loop
Your CEO says, “We need AI.” Before you hire an AI engineer or buy another tool, find one painful process worth fixing.
Episode 15 is our first Dear Human in the Loop, an advice episode for founders, engineering leaders, developers, and decision-makers trying to make practical AI calls.
Dear Human in the Loop
This week’s questions from the inbox.
- Q1. My CEO said, “We need AI.” Where do I start?
Start with problems, not prompts.
- Q2. Should an AI engineer be my startup's first technical hire?
Hire for your biggest business problem, not the trendiest job title.
- Q3. My developers do not trust AI. Should I push them to use it?
Show developers real value and give them time to learn. Trust beats mandates.
- Q4. Which skills will still matter five years from now?
Learn engineering fundamentals, system design, product thinking, communication, and critical thinking. Model quirks will expire.
Sources referenced
Previously
Apple Sues OpenAI and Meta Ships Muse Spark 1.1
EP 14Jul 14, 2026 · 92 minNo Jargon Required
Apple says OpenAI built its hardware program with stolen trade secrets. OpenAI denies any interest in Apple's secrets. We make the Signal or Noise call.
Signal or Noise
The week’s AI headlines, filtered.
- SIGNALApple sues OpenAI over alleged trade-secret theft
- SIGNALMeta releases Muse Spark 1.1 and opens the Meta Model API to developers
- SIGNALOpenAI releases GPT-5.6 across ChatGPT, Codex, and the API
- SIGNALSpaceXAI releases Grok 4.5 for coding and agentic work
- SIGNALAnthropic and OpenAI trade usage resets during the GPT-5.6 launch
No Jargon Required
we explain two terms a decision-maker needs this week.
Multi-agent orchestration: when several AI workers help and when they multiply failure
Model routing: how a product switches among fast, cheap, and capable models
Hot takes
Two opinions, no disclaimers.
Oscar“The Apple lawsuit is a warning for every AI company hiring from a competitor. Talent does not arrive empty-handed. If your onboarding process does not separate experience from confidential material, your product roadmap can become evidence.”
Matt“I don't care if your model is as good as Fable. If it's close, and it's faster, and it's cheaper — it wins. The whole industry is selling you a god-tier model like it's the finish line, but with orchestration and model routing, you don't need one genius doing everything. You need a smart conductor and a cheap, fast bench. The frontier model is becoming the part you use least — and the teams still paying premium tokens for every keystroke are going to feel really dumb in a year.”
Previously
Meta Compute, OpenAI Equity, and AI Startup Bets
EP 13Jul 7, 2026 · 73 minShip It or Skip It
Meta spent so much on AI that it may have backed into becoming a cloud company. Smart move, or capex panic? We make the call.
Welcome to Human In the Loop, a weekly podcast where two builders talk about what matters in AI. No hype. No doomerism. No recaps for the sake of recaps.
Signal or Noise
The week’s AI headlines, filtered.
- SIGNALMeta building a cloud business for excess AI compute
- SIGNALOpenAI reportedly proposing to give 5% equity to a U.S. sovereign wealth fund
- SIGNALClaude Sonnet 5 shipping across Claude, Claude Code, and the Claude Platform, which we call noise until it changes builder behavior
- SIGNALTogether AI raising $800M for open-source AI infrastructure
- SIGNALClaude Science launching as a research workbench
Ship It or Skip It
we debate two AI business ideas.
Meta Compute broker for AI teams
Back-office agent for small law and accounting firms
Hot takes
Two opinions, no disclaimers.
Oscar“The model race is becoming an infrastructure race again.”
Matt“Most AI workbench products are selling relief from tool chaos.”
Previously
The Frontier Just Got a Guest List
EP 12Jun 30, 2026 · 106 minNo Jargon Required
OpenAI just shipped its most powerful model to twenty companies the government picked. Not the best twenty. The approved twenty. A week after the US switched off Anthropic's best model, the frontier has a guest list, and you are not on it. We figure out what that does to anyone trying to build on the newest model.
Welcome to Human In the Loop, a weekly podcast where two builders cut through the AI hype cycle. No breathless hype. No doomerism. No surface-level recaps.
Signal or Noise
The week’s AI headlines, filtered.
- SIGNALGPT-5.6 ships under US government restrictions
OpenAI began a limited preview of GPT-5.6 — Sol, Terra, and Luna — to roughly twenty pre-approved organizations at the US government's request, the first major model to ship under the June 2 executive order. Last week Anthropic went dark by force. This week OpenAI went gated by choice. Two labs, two weeks, same direction.
- SIGNALAnthropic accuses Alibaba's Qwen of the largest distillation attack on record
Anthropic's complaint to Senators Warren and Scott alleges Alibaba's Qwen lab ran about 25,000 fraudulent accounts and 28.8 million exchanges against Claude between April 22 and June 5, aimed at software engineering and agentic reasoning. That is roughly 1.7x the combined total it attributed to DeepSeek, Moonshot, and MiniMax in February, and the first time it has named a major Chinese conglomerate.
- SIGNALOpenAI and Broadcom unveil Jalapeño, OpenAI's first inference chip
OpenAI's first custom chip, built with Broadcom for inference, went from first design to tape-out in nine months with help from OpenAI's own models — the companies call it the fastest ASIC cycle ever. Early claims cite performance-per-watt well above the current state of the art, with first deployment targeted for the end of 2026.
- SIGNALAgentjacking: the prompt-injection hole in every coding agent
A new attack class hijacks coding agents like Claude Code, Cursor, and Codex by hiding instructions inside data they already trust, such as a Sentry error report pulled in over MCP. There is no universal patch — the fix is to treat incoming data as untrusted and keep a human review step before the agent acts. OWASP says prompt injection is up 340% year over year.
- NOISEChatGPT slips below 50% as Gemini and Claude surge
Reports put ChatGPT at 46.4% of the assistant market, its first time below 50% since launch, with Gemini at 27.7% and Claude at 10.3%. Mostly a distribution story — Gemini's climb is Android OS-level placement. Noise for builders, except the one real line: Claude roughly quadrupled its monthly users in five months on the back of agentic coding.
No Jargon Required
Two AI words an exec heard this week and couldn't define on the spot, in plain English.
Distillation
Teaching a small, cheap model by having it copy a big, expensive one's answers. The word at the center of the Alibaba–Claude fight, and why some cheap models punch far above their price. The real question is provenance, not the technique.
Prompt injection
A model reads instructions and data as one stream of words, so commands hidden in the data get obeyed like orders. Researchers say it may not be patchable — the defense is design: read-only by default, a human in front of anything you can't undo.
Hot takes
Two opinions, no disclaimers.
Oscar“The frontier is being quietly nationalized, and most builders are cheering it as safety. Build on the model you can actually keep.”
Matt“Everyone panicking about distillation forgot the lesson. The frontier has no moat. Stop betting your company on a six-week lead anyone can clone.”
Sources referenced
- Axios, OpenAI releases GPT-5.6 (Sol, Terra, Luna) under restrictions
- OpenAI, Previewing GPT-5.6 Sol
- Thurrott, OpenAI launches GPT-5.6 in limited preview
- White House, Promoting Advanced AI Innovation and Security
- CNBC, Anthropic accuses Alibaba of campaign to 'illicitly' extract AI capabilities
- The Next Web, Anthropic accuses Alibaba of largest distillation campaign against Claude
- Mobile World Live, Anthropic accuses Alibaba of large-scale distillation attack
- CNBC, OpenAI and Broadcom reveal Jalapeño, first AI chip in partnership
- TechCrunch, OpenAI unveils its first custom chip, built by Broadcom
- OpenAI, OpenAI and Broadcom unveil LLM-optimized inference chip
- Tom's Hardware, Jalapeño is a reticle-sized ASIC built in nine months
- Digital Applied, Agentjacking: your AI coding agent can be hijacked
- Help Net Security, Prompt injection still drives most agentic AI security failures
- TechTimes, Prompt injection may be a permanent flaw, not a patchable bug
- Unit 42 (Palo Alto), Web-based indirect prompt injection in the wild
- TechTimes, ChatGPT's market share falls below 50% for the first time
- Business Standard, ChatGPT market share slips below 50% as Gemini, Claude gain
- Momentic, Top generative AI chatbots by market share, June 2026
Previously
The World Cup Is Secretly Run by AI
EP 11Jun 23, 2026 · 104 minShip It or Skip It
Last week the US government gave Anthropic's best model a 72-hour public life, then switched it off for every foreign national on earth. This week, three Chinese labs shipped models that beat almost everything in the open, handed over the weights, and charged a tenth of the price. We figure out which of those two plays ends with the whole world building on your stack.
Welcome to Human In the Loop, a weekly podcast where two builders cut through the AI hype cycle. No breathless hype. No doomerism. No surface-level recaps.
Signal or Noise
The week’s AI headlines, filtered.
- SIGNALFIFA is running the 2026 World Cup on AI, the biggest live deployment on earth
An Intelligent Command Center in Miami tied to digital twins of all 16 stadiums, Football AI Pro built with Lenovo on FIFA's own Football Language Model and handed to all 48 teams, 1,248 player avatars from one-second scans, and Gemini drawing up tactics for Argentina, Brazil, and France. Multi-vendor, multi-model, real-time, under a five-second latency budget — the reference architecture nobody asked for.
- SIGNALChina's open-weight wave: GLM-5.2 and Kimi K2.7 Code land
Z.ai's GLM-5.2 tops the open-weight leaderboard with a 1M-token window and MIT weights; Moonshot's Kimi K2.7 Code ships open and tuned for agentic coding. Chinese labs now hold four of the top five open-weight spots at a fraction of Western pricing. The model you self-host is the only one Commerce can't switch off.
- SIGNALAnthropic's Fable 5 and Mythos 5 are still dark a week after the export ban
A full week after the June 12 Commerce directive, both models stay suspended for foreign nationals with no end date. Anthropic published its defense on June 16, arguing a recall over a narrow jailbreak would halt every frontier deployment. This was the closed-frontier fire drill, and most teams failed it.
- SIGNALNoam Shazeer leaves Google for OpenAI
The co-author of "Attention Is All You Need" is gone again, two years after Google paid a reported $2.7B to bring him back into DeepMind. When models converge, talent is the only moat left — and the gravity just pulled toward OpenAI.
- NOISEThe International AI Safety Report 2026 drops
A government-commissioned synthesis meant to set a shared baseline regulators reference, landing the same month the EU AI Act becomes fully applicable. It doesn't regulate anything — it's the document the regulations will point at. Noise for builders, for now.
Ship It or Skip It
Two real ideas from this year's agent discourse, steelman versus pushback.
AI governance layer for the agents your employees already run
The referee between an employee's agent and the action — block the refund agent pushing $40k against a $500 policy before it fires. Own the rules engine, not the dashboard.
Vertical AI harnesses, one trade at a time
Stop selling a chat box, sell the opinionated rig for one trade like marketing or video. Brief in, on-brand output out. Win the domain by being the tool people open every morning, not the model underneath.
Hot takes
Two opinions, no disclaimers.
Oscar“The most important AI lab of 2026 might not be American, and most US builders are too proud to notice.”
Matt“A free model is only free if you own the GPUs, the ops team, and the eval harness to babysit it.”
Sources referenced
- PCMag, inside FIFA's secret AI command center
- Lenovo, AI solutions to power FIFA World Cup 2026
- AkitaOnRails, Kimi 2.7 Code, GLM 5.2, MiniMax M3 local benchmark
- LLM-Stats, Kimi K2.7 Code benchmarks and pricing
- Anthropic, statement on the Fable 5 and Mythos 5 suspension
- Fortune, Anthropic disables Fable and Mythos after US bars foreign access
- CNBC, Google Gemini co-lead Noam Shazeer leaves for OpenAI
- AI Governance Institute, weekly roundup (Safety Report 2026)
- Microsoft, introducing the Agent Governance Toolkit
- Cloud Security Alliance, from guardrails to governance
- GeekWire, the rise of vertical AI agents
- Zapier, the best AI video generators in 2026
Previously
Anthropic's Best Model Got Banned in 72 Hours
EP 10Jun 16, 2026 · 66 minStack Check
Anthropic shipped the most powerful AI ever released to the public on June 9. By June 12 the US government made them pull it back from every foreign national on earth, including Anthropic's own engineers. The most capable model on the planet had a 72-hour public life. We unpack who actually won that week, because it was not the people who got the model.
Welcome to Human In the Loop, a weekly podcast where two builders cut through the AI hype cycle. No breathless hype. No doomerism. No surface-level recaps.
Signal or Noise
The week’s AI headlines, filtered.
- SIGNALAnthropic releases Claude Fable 5 and Mythos 5 at half the old frontier price, then calls for a global AI pause five days earlier. We read the contradiction.
- SIGNALThe Trump administration places Fable 5 and Mythos 5 under export controls after a company jailbreaks Mythos. Anthropic disables access for all customers worldwide to comply, then disputes the order.
- NOISEOpenAI files confidentially for its IPO at a reported $730 to $850 billion, one week after Anthropic filed at $965 billion. Two unprofitable labs, public markets, same week.
- SIGNALOpenAI wires its models and Codex into Oracle Universal Credits, deleting the procurement step for every Oracle enterprise customer.
Stack Check
AI agents in production: the three layers you actually need to run an agent unattended.
Claude Agent SDK
The agent loop you do not have to write.
LangGraph plus LangSmith
Orchestration you can audit.
Temporal
Durable execution so a crash is not a restart.
Hot takes
Two opinions, no disclaimers.
Oscar“The Mythos ban is the best thing that ever happened to open weights. If the Commerce Department can switch off your stack on a Friday, you are renting, not building.”
Matt“Everyone watched the model and missed the money. OpenAI spent the week removing every reason a CFO can say no.”
Sources referenced
Previously
The Government Gets Your AI 30 Days Early + The Token Bill
EP 09Jun 9, 2026 · 99 minNo Jargon Required
The White House just signed an order that lets the federal government test the most powerful AI models up to 30 days before you can touch them. The labs said yes. Real national security, or the backdoor frontier labs were begging for? We make the call.
Signal or Noise
The week’s AI headlines, filtered.
- SIGNALTrump's frontier-model executive order, a voluntary 30-day government early-access window, with the NSA deciding which models qualify
- SIGNALGitHub Copilot flips to token billing on June 1, and developers report bills jumping 10x to 50x
- SIGNALMicrosoft's MAI-Code-1-Flash and Google's $100 dev tier crash the AI coding party Anthropic has owned
- SIGNALChatGPT crosses 1 billion monthly users, and why the real story is Claude growing 640%
- SIGNALAnthropic calls for a global pause on frontier AI, and the 80% of its own code now written by Claude
No Jargon Required
the three words on every AI invoice, explained for anyone who signs off on the bill.
Tokens, and why your AI bill went usage-based
Inference vs training, and what "inference-efficient" actually buys you
Evals and benchmarks, and why the leaderboard is lying to you
Hot takes
Two opinions, no disclaimers.
Oscar“The executive order is not about safety. It is a distribution deal with a 30-day government waiting room on every frontier release.”
Matt“The Copilot billing meltdown is the best thing to happen to engineering leaders this year. The meter was always running. Now you can see it.”
Previously
Anthropic's "Too Dangerous" Model Is Coming for Everyone
EP 08Jun 2, 2026 · 123 minShip It or Skip It
Six weeks ago Anthropic said a model was too dangerous to ship. This week they shipped Opus 4.8, called it near-Mythos, and said Mythos-class models reach every customer in the coming weeks. Was it ever that dangerous? We make the call.
Signal or Noise
The week’s AI headlines, filtered.
- SIGNALClaude Opus 4.8, near-Mythos alignment, 3x cheaper fast mode, and Dynamic Workflows that fan one model across up to 1,000 subagents
- SIGNALAnthropic passes OpenAI, a new round values it near $1 trillion, plus a projected first operating profit on $10.9B revenue
- SIGNALCognition raises $1B at $26B, Devin's autonomous coding agent hits a ~$492M run rate
- SIGNALA poisoned Nx Console VS Code extension stole Claude Code configs and helped clone 3,800 of GitHub's internal repos
- SIGNALPope Leo XIV's first encyclical takes on AI, signal for regulated buyers, noise for your standup
Ship It or Skip It
three real ideas from this month's builder discourse.
The Company Brain
My First Million Ep 822
Voice AI agent for the trades
Avoca's $125M raise
Stripe for AI agents
Circle Agent Stack
Hot takes
Two opinions, no disclaimers.
Oscar“Too dangerous to ship" was never a safety call. It was a 90-day enterprise head start.”
Matt“Devin at $26B is the top, not the floor. An agent lab with no distribution is the WeWork of 2026.”
Previously
Cursor's $0.50 Coding Model + Anthropic's $900B Round
EP 07May 26, 2026 · 98 minStack Check
Cursor just shipped a coding model that matches Claude Opus 4.7 at one tenth the price, built on an open Chinese base model. The same week, Anthropic is closing a $30B round at a $900B valuation. If frontier coding got commoditized in seven days, what is anyone paying premium for?
Welcome to Human In the Loop, a weekly podcast where two builders cut through the AI hype cycle. No breathless hype. No doomerism. No surface-level recaps. Honest, opinionated conversations from people actually building with AI.
Signal or Noise
The week’s AI headlines, filtered.
- SIGNALCursor Composer 2.5, top-three coding agent at one tenth the cost of Opus
- SIGNALAndrej Karpathy joins Anthropic's pre-training team
- SIGNALAnthropic in talks for a $30B round at a $900B valuation
- SIGNALGoogle I/O 2026, Gemini Spark agent and Antigravity 2.0
- SIGNALHark's $700M Series A for a "secret" AI device, why we are calling it noise for now
Stack Check
three AI dev tools we actually adopted in the last 30 days. What each one replaces, why it matters, and would we actually keep using it.
Matt Pocock's `skills` repo, Skill Handoff
a real replacement for `/compact`
Convex (convex.dev), TypeScript-native backend built for agentic workflows
Google Antigravity 2.0 + CLI + Managed Agents API
Hot takes
Two opinions, no disclaimers.
Oscar“Cursor Composer 2.5 just proved the open base model thesis. Coding-as-a-service is about to fall off a cliff in price.”
Matt“Karpathy joining Anthropic is worth more than the $900B valuation as a signal. Talent is the only moat the market still underprices.”
Previously
Anthropic Killed the AI-for-SMB Playbook + AI Agents 101
EP 06May 19, 2026 · 92 minNo Jargon Required
Three months ago on Episode 1, we asked if there was a real business in selling AI to small businesses. This week, Anthropic shipped it themselves. So we just watched the founder playbook get eaten by a frontier lab in 90 days. If you were six weeks into building that company, what do you do Monday?
Signal or Noise
The week’s AI headlines, filtered.
- SIGNALAnthropic ships Claude for Small Business
15 agentic workflows, QuickBooks + HubSpot + Microsoft 365 connectors, 10-city US tour.
- SIGNALThinking Machines previews “Interaction Models”
276B MoE, 0.4s full-duplex response, Mira Murati throws shade at every other lab's voice stack.
- SIGNALSierra raises $950M at $15B+ valuation
$150M ARR, 40% of the Fortune 50 paying for customer agents.
- SIGNALTrump administration reverses on AI oversight
Driven by national-security fears around Anthropic's Mythos model.
- SIGNALCloudflare cuts 1,100 jobs and blames AI
On the same morning it posts record revenue — the first big “AI did it” layoff that isn't a cover for bad numbers.
No Jargon Required
Three AI concepts every exec and operator keeps hearing this quarter, in plain English.
AI Agent
The word every vendor abuses — and how to tell the real ones from the chatbots in costume.
Connectors and MCP
What “AI plugged into your tools” actually requires.
Full-Duplex Voice (Interaction Models)
What to ask when a vendor pitches a voice agent.
Hot takes
Two opinions, no disclaimers.
Oscar“Anthropic just turned every AI consultant into a reseller. By Q3, the model providers own the workflow layer for every regulated and semi-regulated vertical.”
Matt“Every vendor calling their thing an ‘AI agent’ in 2026 is going to look like every vendor that called their thing ‘cloud-native’ in 2015. By Q4 the term means nothing.”
Sources referenced
- Anthropic — Claude for Small Business
- TechCrunch — Anthropic SMB launch
- TechCrunch — Sierra raises $950M
- TechCrunch — Thinking Machines Interaction Models
- Fortune — Trump admin embraces AI oversight
- Cloudflare — Building for the future (Matthew Prince memo)
- TechCrunch — Cloudflare says AI made 1,100 jobs obsolete
- Anthropic — Model Context Protocol
- Notion — Developer Platform launch
Previously
Self-Hosted AI with Jackson Oaks. Why 80% of AI Bills Are Wasted.
EP 05May 12, 2026 · 58 minGuest Deep-Dive
On May 12, 2026, Jackson Oaks — founder of Recursion AI and the self-hosted AI platform Courier — joins the show. Jackson has been running open-source models in production for SMBs since early 2025, back when most of the industry still called local inference a hobbyist toy. Today his company runs production workloads on Apple Mac hardware, processes 600 to 800 thousand API calls a month per machine, and is profitable doing it.
His thesis: eighty percent of businesses have eighty percent of use cases that don't need Opus or GPT-5.5. They need a smaller open-source model, good systems, and predictable pricing. Frontier intelligence, or boring infrastructure? Per-token billing, or a flat rate on hardware you already own?
That's the cold open. Every fifth episode of Human In The Loop is a guest deep-dive — no Signal or Noise, no rotating segment, one topic, one expert. Oscar, Matt, and Jackson spend 58 minutes on what most AI bills are actually paying for, and close with three hot takes you won't hear anywhere else.
Guest Deep-Dive
What we cover with Jackson Oaks.
The 80/20 of open-source AI
Where local models win, where frontier models still matter, and how to tell which category your use case is in.
Three measurable ROI categories
Time bought back, margin protected on AI products, conversion lift on AI features — and which one is nearly impossible to measure honestly.
The Apple long-bet
Why M-series is the most underrated AI hardware story of the decade. 1/10th the cost of NVIDIA, 30 to 40x more power efficient. Apple has been positioning for this since the M1 dropped — the same week ChatGPT did.
The economics
600 to 800,000 API calls a month on a single Mac Studio. Flat-rate cloud pricing starting at $100/mo for 45 to 100K calls. The math on why predictable beats per-token for SMBs.
The Deloitte story
A Fortune client paid Deloitte $400,000 for what turned out to be a GPT wrapper with an unoptimized RAG database. MIT says 95% of business AI pilots can't measure ROI. We have theories.
War stories
The infinite hallucination loop that ate 10,000 requests in a week. Building hallucination detection from scratch. The fine-tuned 14B model that outperformed GPT-4o on a real production task because the data was good and the task was narrow.
The boring stuff nobody talks about
Idempotency in non-deterministic systems. Observability. Redundancy. Why your AI feature needs the manual OCR backup you were about to delete.
The vertical opportunity
Why the next wave of AI businesses isn't “AI for everything” — it's $200/mo agents for pressure washers, plumbers, landscapers, attorneys, accountants. Specialize down, charge real money, capture real value.
Hot takes
Two opinions, no disclaimers.
Matt“One of Anthropic or OpenAI buys the other — or merges with a foreign lab — inside three years. Or open-source overtakes both and they shrink to a fraction of their current valuations. Their burn rate is not sustainable.”
Oscar“Every home will eventually have a Mac Mini running a better Alexa, powered by open-source models — not closed-source ones. That future is closer than people think.”
Jackson“Better systems and structure with smaller open-source models beat throwing frontier models at unstructured problems. People are compensating for a lack of system with a bigger model. We're doing the opposite — we're building better systems so we can use smaller models.”
Previously
9 Seconds to Delete a Database. Anthropic at $900B.
EP 04May 5, 2026 · 46 minShip It or Skip It
On April 24, 2026, an AI coding agent hit a credential mismatch in a staging environment. Nine seconds later, the entire production database of a company called PocketOS was gone. Backups too. The agent confessed: “I violated every principle I was given.” The model was Claude Opus 4.6.
Five days later, on April 29, Bloomberg and CNBC reported Anthropic was weighing offers at a $900 billion valuation. Roughly $50 billion in new capital. A round that would put the lab behind Claude above OpenAI as the most valuable AI company in the world. Two stories. Same week. Same lab. We call every one.
That's the cold open. From there, Oscar and Matt spend 46 minutes doing what this show does: filter the week's AI noise, debate three real business ideas, and close with two hot takes you won't hear anywhere else.
Signal or Noise
The week’s AI headlines, filtered.
- SIGNALAnthropic at $900B
$50B raise on the table. Run-rate revenue at $30B. The Mythos withholding story we called in Episode 1 was the brand. The Google $40B in Episode 3 was the funding story. This is the receipt. Treat your model spend like an interest rate, not a software line item.
- SIGNALAI agent deletes PocketOS database
Nine seconds. Backups gone. The agent went looking for an API token, found one in an unrelated file, and used it. The token was scoped for any operation, including destructive ones. Three holes in the swiss cheese. The agent walked through all of them.
- SIGNALChina blocks Meta–Manus, Meta buys ARI
Beijing killed Meta's $2B Singapore-routed AI deal on April 27. Meta acquired humanoid-robotics startup Assured Robot Intelligence on May 1. The thesis didn't change. The substrate did. Robotics is the next frontier-model fight.
- SIGNALCloudflare and Stripe ship agent provisioning
Agents can now create accounts, register domains, start paid subscriptions, and deploy apps with no human in the loop. Default spend cap is $100 per provider per month. The PocketOS story is the demand signal for spend caps and audit logs.
- SIGNALChinese court rules AI-replacement layoff illegal
The Hangzhou Intermediate People's Court upheld a ruling that Zhou, a QA supervisor verifying LLM outputs, couldn't be fired or pushed to a 40% pay cut just because AI took over his work. First major court ruling anywhere putting limits on AI-driven layoffs.
Ship It or Skip It
Three real ideas from YC's RFS, no fence-sitting.
AI-Native Service Companies
Per Gustaf Alströmer's RFS. Don't sell SaaS to insurance brokers, accountants, or compliance teams. Become the broker. File the taxes. Run the audit. Charge services prices for AI-margin work. The wedge is bigger than the entire SaaS market. Pushback: you stop being software and start being ops, with E&O liability and license requirements baked in.
Company Brain
Per Tom Blomfield's RFS. The blocker to AI automation isn't the models anymore. It's domain knowledge scattered across heads, email, Slack, tickets, and databases. Build the layer that pulls fragmented company knowledge, keeps it current, and turns it into an executable skills file for agents. Pushback: this is wikis 5.0, and Microsoft and Notion eat the category as a feature in 18 months.
Counter-Swarm Defense
Per Tyler Bosmeny's RFS. $500K interceptors don't work against $500 drones. Build the Cloudflare-shaped layer for swarm defense: distributed sensors, software-first interceptors, autonomy-stack attacks. The thesis is real. The dual-use moral question is not optional.
Hot takes
Two opinions, no disclaimers.
Oscar“Anthropic at nine hundred billion dollars is the moment the labs stopped being labs. They're sovereign wealth funds with a research division attached. The Mythos withholding story we called in Episode 1 was the brand. The fundraise is the receipt. Price your model spend like an interest rate, not a software line item.”
Matt“Every AI agent in production is one bad scope away from being the PocketOS story. The fix isn't better models. It's permissions, audit logs, and a human approval gate on every destructive operation. If your agent can drop a database, your agent will drop a database. The question isn't if. The question is on what week.”
Sources referenced
- Bloomberg — Anthropic at $900B valuation
- CNBC — Anthropic $900B fundraise
- TechCrunch — Anthropic $50B round
- The Register — Cursor-Opus agent deletes PocketOS database
- Tom's Hardware — 9 second deletion
- TechCrunch — China blocks Meta–Manus
- Bloomberg — Meta acquires Assured Robot Intelligence
- Cloudflare — Agents can buy domains and deploy
- NPR — Chinese tech worker laid off and replaced by AI
- Latent Space — Every Agent Needs a Box (Aaron Levie)
Previously
GPT-5.5 Beat Mythos. The Race Got Weirder.
EP 03Apr 28, 2026 · 49 minStack Check
On April 23, 2026, OpenAI shipped GPT-5.5, the model previously known as Spud. The same day, VentureBeat clocked it narrowly beating Anthropic's gated Claude Mythos Preview on Terminal-Bench 2.0. The day after, DeepSeek released V4, a 1.6 trillion parameter open-source model running natively on Huawei chips, at one-sixth the inference cost of the closed frontier. And Google announced it would invest up to forty billion dollars in Anthropic.
Two weeks ago we called GPT-5.5 noise. Now it's the lead story. The frontier is moving every six weeks, the cloud war is inside the labs, and the open-source ceiling jumped a tier on hardware your security team can't audit.
That's the cold open. From there, Oscar and Matt spend 49 minutes doing what this show does: filter the week's AI noise, review the tools they actually use, and close with two hot takes you won't hear anywhere else.
Signal or Noise
The week’s AI headlines, filtered.
- SIGNALGPT-5.5 (Spud) ships
The first buyable frontier model to beat a gated frontier model on a public benchmark. Same week as the Google–Anthropic announcement. The two-tier story from Episode 2 just collided with the cadence story Fortune was writing all month.
- SIGNALDeepSeek V4
1.6T total parameters, 1M context, MIT license, native on Huawei Ascend 950PR. The open-weights ceiling moved by a tier. Geopolitics moved into the model card.
- SIGNALGoogle + Anthropic, up to $40B
Google has Gemini in-house and is still writing a forty billion dollar check. That tells you what they actually believe about the next 18 months.
- SIGNALCognition at $25B
Devin-in-Windsurf is the first time the agent and the IDE feel like one product, not two stitched together. Whether the price tag holds is a different question.
- NOISEGrok 4.3 Beta at $300/month
Video and slide generation, no cross-session memory. The price tag is the message.
Stack Check
Four tools we actually use.
Vercel Labs Agent Browser
Browser automation CLI built for agents. Persistent sessions, headless or real Chrome with your profile, annotated screenshots so a model can click “label 7” instead of guessing CSS selectors.
WorkOS (FGA + AuthKit)
The auth and permissions layer for agent surfaces. Sub-50ms p95 authorization checks. AI Installer in the CLI sets up auth in under five minutes. Last week's Vercel breach was a permissions story with a price tag attached.
AI SDK (ai-sdk.dev)
Define the agent once, swap the model in one line. The Gateway added GPT-5.5, DeepSeek V4, Kimi K2.6, and GPT Image 2 inside two weeks. The only honest way to evaluate this week's launches on real workloads.
Claude Code /remote-control
Run claude remote-control on your laptop, approve file changes from your phone, walk to lunch. Work stays local. The walkaway-and-approve workflow is the productivity win of the quarter.
Hot takes
Two opinions, no disclaimers.
Oscar“Open-source AI just won the cost war and lost the trust war in the same week. DeepSeek V4 is the best open weights anyone has ever shipped. It also runs on chips your security team can't audit.”
Matt“OpenAI shipping GPT-5.5 two weeks after GPT-5.4 is the end of the model launch as a marketing event. From now on, the model layer is a software update, and the only thing that matters is what you ship on top of it.”
Sources referenced
- OpenAI — Introducing GPT-5.5
- TechCrunch — OpenAI releases GPT-5.5
- Fortune — AI launches look like software updates
- VentureBeat — GPT-5.5 beats Mythos on Terminal-Bench
- TechCrunch — DeepSeek V4 closes the gap with frontier
- Fortune — DeepSeek V4 on Huawei chips
- CNBC — Google to invest up to $40B in Anthropic
- Bloomberg — Cognition at $25B
- vercel-labs/agent-browser on GitHub
- WorkOS — Authorization layer for AI agents
- AI SDK Docs (ai-sdk.dev)
- Claude Code — Remote Control
Previously
The Second-Best Model You Can Actually Buy
EP 02Apr 21, 2026 · 47 minNo Jargon Required
On April 16, 2026, Anthropic shipped Claude Opus 4.7 with a 1M token context window at standard API pricing, a 128k max output, and a new “xhigh” effort level tuned for coding and agentic work. In the same announcement, Anthropic said on record that 4.7 is “less risky than Mythos,” and that Mythos stays gated to roughly 50 enterprise partners under Project Glasswing. So the flagship you can buy is, by Anthropic's own words, their second best.
Radical honesty, or the sharpest upsell in enterprise AI since cloud credits?
That's the cold open. From there, Oscar and Matt spend 45 minutes doing what this show does: filtering the week's AI noise, breaking down the concepts decision-makers actually need, and closing with two hot takes you won't hear anywhere else.
Signal or Noise
The week’s AI headlines, filtered.
- SIGNALClaude Opus 4.7
1M context at standard pricing is the real story. A meaningful win for long-horizon agent workloads. And Anthropic just taught every enterprise buyer to ask, “what's the gated tier.”
- SIGNALOpenAI's updated Agents SDK
Sandboxed agents as a primitive. Python first, TypeScript later. Good news for builders, and great news for OpenAI's runtime.
- SIGNALAI on the desktop
Google's native Gemini macOS app plus Perplexity's Personal Computer. The OS is replacing the browser as the AI UX primitive.
- SIGNALStanford's 2026 AI Index
Only 10% of Americans are more excited than concerned about AI, versus 56% of experts. Your consumer product has a trust problem, not a capability problem.
- NOISEAllbirds pivots to AI
A struggling sneaker company rebrands as NewBird AI, sells the footwear business, drops its environmental charter, and the stock jumps 582%. The “pivot to AI” is the new “pivot to crypto.” Price your term sheets accordingly.
No Jargon Required
Three concepts decision-makers need to call this quarter.
Context Engineering
Prompt engineering is writing a good email. Context engineering is designing the whole office the recipient works in. When every company can access the same foundation models, the wiring is the moat.
Agentic AI
What an agent actually is, and why Gartner says 40% of enterprise agentic AI projects will be canceled by the end of 2027 because of cost, unclear value, and weak risk controls.
Frontier Model Tiering
The top AI labs now ship a public tier and a gated tier. The best model is probably one you cannot buy. How to build strategy around that reality.
Hot takes
Two opinions, no disclaimers.
Oscar“Context engineering isn't a new skill. It's ‘write good internal documentation’ rebranded so consultants can charge for it. The companies winning in 2026 had clean internal data in 2022.”
Matt“Every executive asking ‘what's our agentic AI strategy’ in April 2026 is two years from being replaced by one who already shipped something.”
Sources referenced
- Anthropic — What's new in Claude Opus 4.7
- CNBC — Opus 4.7 less risky than Mythos
- Axios — Anthropic concedes Opus 4.7 trails Mythos
- OpenAI — The next evolution of the Agents SDK
- Google — Gemini for Mac
- IEEE Spectrum — Stanford's 2026 AI Index
- Gartner — 40% of agentic AI projects canceled by 2027
- CBS News — Allbirds ditches footwear and pivots to AI
- Fortune — Allbirds' desperate AI pivot and 600% stock surge
Previously
The Model Anthropic Refused to Ship
EP 01Apr 14, 2026 · 47 minShip It or Skip It
On April 7, 2026, Anthropic unveiled Claude Mythos Preview — a frontier model they describe as a “step change” in capability — and then refused to release it. Instead: Project Glasswing, a gated program for Amazon, Apple, Microsoft, Cisco, CrowdStrike, and a handful of others. The first major withheld frontier model in roughly seven years.
Responsible scaling finally biting? Or enterprise GTM dressed up as safety PR?
That's the cold open. From there, Oscar and Matt spend 45 minutes doing what this show does: filtering the week's AI noise, debating real business ideas, and closing with two hot takes you won't hear anywhere else.
Signal or Noise
The week’s AI headlines, filtered.
- SIGNALClaude Mythos Preview
Anthropic used Mythos to identify thousands of zero-days across every major OS and browser, then chose not to ship. If Mythos-wielding attackers find what small shops can't patch, the defender/attacker asymmetry just broke.
- SIGNALMeta Muse Spark
First model from Meta Superintelligence Labs. Llama-4 midsize quality at ~10x less compute, shipping into Facebook, Instagram, WhatsApp, and Ray-Ban Meta glasses. If you build consumer or wearable UX, unit economics just moved.
- NOISEOpenAI “Spud”
Polymarket gives GPT-5.5/6 a 78% chance of shipping by April 30. Predictions aren't news. The real story is OpenAI's $122B raise.
- SIGNALThe anti-distillation pact
OpenAI, Anthropic, and Google coordinating against Chinese labs. IP defense or three competitors ganging up on a fourth because it's easier than competing on capability?
- SIGNALAnthropic × Coefficient Bio, $400M
Horizontal labs are quietly going vertical. Life sciences. Cybersecurity. Legal next? Finance? Pick your domain before the labs pick it for you.
Ship It or Skip It
Three real ideas, no fence-sitting.
“Bring AI to small businesses”
Via My First Million Ep 811. Productized 30-day AI-ops install for HVAC, dental, property management. Software margins, services delivery. But the real question: are you the next picks-and-shovels play, or just another consultant in a trench coat?
Vertical AI agent for insurance claims
Per-claim pricing instead of per-seat. Eats into labor P&L, not software budget. But insurance = 18-month sales cycles. You die of starvation before your first logo unless you have an insider co-founder.
“Lovable-for-X”
Clone the $200M ARR playbook for a regulated vertical — legal ops, clinical workflows, HR. Compliance moat is real. But is it also the ceiling on product velocity?
Hot takes
Two opinions, no disclaimers.
Oscar“Anthropic withholding Mythos is the new moat. Safety-as-marketing is about to become the dominant frontier-lab playbook — because nobody wants to be the lab that shipped the model that broke the internet.”
Matt“The “AI for small businesses” gold rush is 90% consultants in trench coats. The real money isn't selling AI to SMBs — it's selling picks-and-shovels to the 10,000 consultants who are.”
New episodes every week. Reply with the story you want us on next.
Subscribe