Monday, October 5, 2026
The Death of the AI Vibe
October 5 · 18 videos
Vibes are dead.
Production is the standard.
Data must be clean.
Agents must be verified.
The moat is evaporating.
Build the right thing.
“The transcript looked fine. The call was not.”
From Raw Documents to AI-Ready Data: Leo Platzer and Jeff Koss
Leo Platzer · AI Engineer · 21 min
Watch on YouTube →Leo Platzer and Jeff Koss discuss the transition from small RAG pilots to large scale production systems. They highlight how uncurated enterprise data leads to context crowding and degraded performance.
- Moving from 40 files to 80,000 documents often causes RAG systems to fail due to stale or duplicative data.
- Approximately 30 percent of a typical document corpus is redundant which can pollute up to 80 percent of an agent context window.
- Automated data curation can reduce preparation time from four months to just a few days.
- Cleaning a corpus to remove duplicates can double recall in multi-hop RAG evaluations.
- Providing confidence scores for AI generated metadata helps business users trust the data products.
From Vibes to Production: Evaluating and Shipping AI Agents That Work 101
Laurie Voss · AI Engineer · 111 min
Watch on YouTube →Laurie Voss explains how to move beyond shipping AI on vibes by implementing rigorous evaluation loops. This workshop covers instrumenting agents and building deterministic defense layers for reliable production.
- Human judges are inconsistent and can miss up to 50 percent of defects due to fatigue.
- Eval Driven Development uses production traces to create regression datasets for automated iteration.
- Developers need 200 to 400 samples to make confident shipping decisions for AI features.
- Data quality improvements offer a higher return on investment than prompt iteration or model tuning.
- Defining success is a product requirement that requires domain experts to calibrate automated judges.
The Transcript Looked Fine. The Call Wasn't. : Debugging Voice Agents, Arize
Fuad Ali · AI Engineer · 18 min
Watch on YouTube →Fuad Ali explores the challenges of debugging voice agents where text transcripts fail to capture critical audio issues. He advocates for integrated observability that combines audio and WebSocket events.
- Traditional text logs miss critical failure modes like latency and turn taking issues in voice agents.
- Audio native evaluations use models to judge tonality and sentiment rather than just text.
- Provider agnosticism allows developers to swap model providers without breaking observability schemas.
- Transcripts are lossy and do not show dead air or interruptions that ruin user experience.
- The vision for voice agents includes a self healing loop where SRE agents hypothesize and validate fixes.
Research to Reality with Google DeepMind: Paige Bailey, Google DeepMind
Paige Bailey · AI Engineer · 10 min
Watch on YouTube →Paige Bailey introduces the Gemma 4 model family and the shift toward high performance local inference. She demonstrates how generative AI can now run entirely within browsers and on mobile devices.
- Gemma 4 models range from 2 billion to 31 billion parameters and use Apache 2 licensing.
- The 2 billion parameter model can run on a phone with a footprint of less than 1GB.
- Moving inference to local devices reduces API costs and removes the need for distributed infrastructure.
- Managed agents can execute complex tasks in a Linux sandbox using WebAssembly in the browser.
- High performing small models can replace much larger ones to provide better hardware ROI.
PewDiePie is setting AI free... and OpenAI is furious
PewDiePie · Fireship · 5 min
Watch on YouTube →This video covers PewDiePie's transition into AI engineering and his development of the uncensored Ajax model. It highlights the growing movement toward sovereign intelligence and local model ownership.
- Ajax is an uncensored 9 billion parameter model based on Alibaba Qwen 3.5.
- The project uses Group Relative Policy Optimization to eliminate the need for a separate critic model.
- OpenAI has banned users for distillation which is the process of training small models on larger model outputs.
- The Heretic tool was used to surgically remove obliteration parameters to de-censor the model.
- Sovereign intelligence allows individuals to own their weights rather than renting from corporate providers.
Build the Right Thing: Product Engineering (Part 2) — Kent C. Dodds, EpicProduct.engineer
Kent C. Dodds · AI Engineer · 52 min
Watch on YouTube →Kent C. Dodds explains why product judgment is the primary differentiator for engineers as AI commoditizes software implementation. He uses the Kano Model and Jobs to Be Done framework to show how to identify core user needs over feature requests.
- Engineers must act as the final gatekeepers of product ethics by pushing back on harmful or invasive features.
- The Kano Model categorizes features into basic needs, performance needs, and delighters to prioritize technical effort.
- User error is viewed as a failure of system design rather than a mistake by the individual using the product.
- Success is measured by the Little Hire which occurs every time a customer uses the product to make progress.
- Features naturally migrate over time from being delighters to becoming basic expectations in the competitive landscape.
Build the Right Thing: Product Engineering (Part 1) — Kent C. Dodds, EpicProduct.engineer
Kent C. Dodds · AI Engineer · 55 min
Watch on YouTube →This workshop explores the shift from implementation to product engineering where judgment becomes the most durable skill for developers. Kent C. Dodds introduces frameworks like The Mom Test to validate ideas before investing in expensive technical builds.
- Product engineering requires developers to balance technical implementation with a deep understanding of customer problems.
- The Mom Test focuses on asking users about their past behaviors and current workarounds instead of future speculations.
- Building the smallest useful slice helps teams avoid the sunk cost fallacy and validate products with minimal investment.
- Implementation is becoming a commodity while the real business risk lies in building features that provide no value.
- Validation can often be done in weeks using manual integrations rather than spending months on unproven software.
How I Learned to Stop Worrying and Love the Sandbox — Matt Brockman, E2B
Matt Brockman · AI Engineer · 59 min
Watch on YouTube →Matt Brockman from E2B provides a technical deep dive into managing high performance sandboxes for AI agents. He discusses the operational challenges of scaling isolated environments while maintaining sub 100ms startup times.
- E2B uses memory and filesystem snapshotting to allow agents to resume complex tasks without losing context.
- The Adaptive Timeout Policy grants long execution times for active tasks while using aggressive timeouts for idle ones.
- Observability through logging and monitoring is essential for tracking resource leaks like CPU and disk exhaustion.
- Scaling to thousands of concurrent users requires automated cleanup policies to manage orphaned processes.
- Developers must balance the speed of running sandboxes against the cost savings of paused or snapshotted environments.
Evals in AI: A Deep Dive — Tejas Kumar, IBM
Tejas Kumar · AI Engineer · 59 min
Watch on YouTube →Tejas Kumar outlines a reliability model for AI that combines runtime harnesses with ahead of time evaluations. He details how to build LLM as a Judge architectures to prevent security breaches and ensure production readiness.
- Evaluations function as fuzzy unit tests for non deterministic systems to ensure reliability before deployment.
- The Meta Instagram chatbot exploit highlights the danger of moving fast without proper evaluation and security guards.
- Pointwise evaluations using binary pass fail metrics are preferred over pairwise comparisons to avoid position bias.
- Teams should aim for an 80 to 85 percent agreement threshold between AI judges and human scored golden datasets.
- Continuous traffic skimming allows developers to update evaluation datasets with real world production signals.
Your AI Agent Is Confidently Wrong About Production — Willem Pienaar, Cleric
Willem Pienaar · AI Engineer · 14 min
Watch on YouTube →Willem Pienaar discusses why AI agents struggle with production debugging due to their training for decisiveness over skepticism. He introduces a grounding framework designed to help agents navigate complex and distributed service failures.
- RLHF often optimizes for confident answers which conflicts with the doubt and uncertainty required for effective SRE.
- The grounding framework uses parallel theory generation and log clustering to identify the root causes of incidents.
- Verification of a fix through a return to normal behavior is the only true metric for agent success in production.
- Trust is built by calibrating agent confidence scores so they accurately correlate with the probability of being correct.
- Production environments lack the immediate feedback loops of development making them significantly harder for agents to master.
Lifestyles of the AI-Native: Nick Nisi and Zack Proser, WorkOS
Nick Nisi · AI Engineer · 61 min
Watch on YouTube →Modern AI engineering requires managing a fleet of concurrent agents rather than supervising a single session. Nick Nisi and Zack Proser demonstrate how a two person team uses voice coding and automated verification gates to manage dozens of repositories.
- Transition from manual input to systems architecture by operating a fleet of twelve concurrent agents via the Fleet tool.
- Increase input bandwidth by using voice coding with Whisper to reach 190 words per minute compared to 90 words per minute typing.
- Implement adversarial reviews where a secondary model like Codeex critiques the primary agent work to ensure quality.
- Use verification gates to prevent agents from faking success such as touching a file to mimic a passing test.
- Adopt progressive disclosure for agent memory by only loading relevant technical context based on the current file focus.
- Focus on fixing the system that generates the code rather than manually fixing the generated code itself.
Agentic Search vs Vector Search for Coding Agents: We Ran the Eval
Jess Wang · AI Engineer · 18 min
Watch on YouTube →This presentation argues that AI development must move from intuition based shipping to systematic evaluations that drive continuous improvement. Jess Wang compares vector search and agentic search for coding tasks to reveal significant cost and efficiency differences.
- Vector search proved to be four times more expensive than agentic search while achieving the same accuracy in bug fixing tasks.
- Traditional RAG based vector search often fails in codebases because it returns chunks missing critical context like imports or function hierarchies.
- The evaluation flywheel transforms production logs into datasets for testing to prevent regressions and improve code quality.
- Effective evaluations require a four part framework consisting of a dataset, task, scorer, and experiments.
- Shipping decisions based on engineering intuition or vibes lead to quantifiable regressions that systematic evals can prevent.
How Founders Win at Fundraising, Sales, and Hiring
Dalton Caldwell · Dalton and Michael · 24 min
Watch on YouTube →Startup success is driven by leverage and the ability to walk away rather than tactical pitch deck optimizations. Dalton Caldwell and Michael Seibel explain how rapid growth shifts the dynamic from founders asking for favors to investors chasing opportunities.
- Top tier venture capitalists actively chase between 80 percent and 90 percent of the deals they eventually fund.
- High leverage is built through years of pre-work and traction rather than the final seconds of a pitch or negotiation.
- Founders must overcome the school mindset of expecting help and instead focus on offering objective value to partners.
- A strong sales position requires a robust pipeline where no single deal has the power to kill the company.
- Direct to user storytelling is a high leverage activity while traditional press coverage rarely drives meaningful signups.
Agent Memory Is Solved. Agent Learning Isn't.: Karthik Ranganathan, Yugabyte
Karthik Ranganathan · AI Engineer · 20 min
Watch on YouTube →While individual agent memory is largely solved, collective agent learning remains a bottleneck for production environments. Karthik Ranganathan discusses how to manage governed and traceable context to prevent token waste and context rot.
- Increasing context windows often leads to expensive failure because context rots and introduces noise over time.
- The Data Pack framework allows agents to push private memories that are then promoted by humans into a shared knowledge pool.
- Iterative tuning of the RAG pipeline increased faithfulness from 14 percent to 82 percent while reducing context size by seven times.
- Capturing the underlying reasoning and dead ends of agent tasks prevents subsequent agents from burning tokens re-deriving information.
- Human promotion of agent memory acts as a high quality training signal for future autonomous orchestration systems.
the paradox of being ambitious but lazy
Rob Dial · The Mindset Mentor Podcast · 17 min
Watch on YouTube →Laziness is a manageable behavior rather than a fixed trait and can be overcome through environmental controls and structured systems. Rob Dial provides a framework for transitioning from unreliable willpower to process driven productivity.
- Reframe procrastination as avoidant behavior to identify the specific tasks or emotions being escaped.
- Break down large financial goals into specific daily activities such as making 50 phone calls to reach a six figure income.
- Use environmental de-stimulation to remove distractions and force the brain to focus on high priority tasks.
- Train focus and productivity as literal muscles through repetition and techniques like the Pomodoro method.
- High stakes motivation can reveal untapped creative income streams when individuals are forced to solve urgent problems.
Bringing Extinct Species Back to Life | Dr. Beth Shapiro
Dr. Beth Shapiro · Andrew Huberman · 135 min
Watch on YouTube →Dr. Beth Shapiro explains how de-extinction technology serves as a vital toolkit for modern conservation and genetic rescue. This discussion highlights the shift from viewing species as static categories to functional biological interactions.
- De-extinction efforts for mammoths and dodos fund the development of critical tools to save endangered species like the black-footed ferret.
- Scientists are using multiplex genome engineering to bridge the one percent genetic gap between Asian elephants and woolly mammoths.
- The project aims to create functional proxies such as gray wolves edited with ancient traits for size and coat color.
- Approximately 95 percent of the Neanderthal genome still exists across the modern human population today.
- Public trust is maintained through transparency measures like Care Reports and independent advisory panels for controversial projects.
- The human-driven rate of environmental change now necessitates intentional genetic intervention to ensure planetary resilience.
The Current State of Consumer AI
Olivia Moore · a16z · 49 min
Watch on YouTube →This analysis explores the shift in consumer AI from simple productivity tools to high-value personal agents and entertainment platforms. It examines why the current subscription model may be a temporary phase before a return to ad-based and transactional revenue.
- Only 4.5 percent of US consumers currently pay for AI subscriptions while the top 1 percent of spenders average over 900 dollars monthly.
- OpenAI has reached a 1 billion dollar annualized advertising run rate as the industry pivots toward traditional monetization models.
- Success in the consumer market is moving toward spend time categories like dating, shopping, and social networking.
- Personal agents face a significant adoption wall due to user privacy concerns regarding access to sensitive email and financial data.
- AI startups are expanding into enterprise and small business use cases much faster than previous software generations.
- Well-designed interfaces benefit from a rising tide effect where model improvements instantly increase the utility of the software layer.
The Billion Dollar AI Advantage Is Disappearing
Dr. Koa · Two Minute Papers · 3 min
Watch on YouTube →Dr. Koa discusses how the competitive moat of massive compute is shrinking as frontier-level intelligence is compressed into smaller and cheaper models. This democratization allows solo developers to achieve results that previously required billion-dollar engineering teams.
- Claude Sonnet 5.5 demonstrates that high-level intelligence is now a function of training quality rather than raw parameter volume.
- New models are achieving state-of-the-art performance at costs 2.5 to 5 times lower than previous benchmarks within weeks of release.
- The industry is entering an era of portable intelligence where frontier models can run locally on consumer laptops and smartphones.
- Complex physics simulations that once required massive infrastructure can now be ported into single HTML files using efficient models.
- The rapid narrowing of release windows for cheaper successors forces developers to prioritize adaptability over specific model loyalty.
References
PeopleLeo Platzer · Jeff Koss · Laurie Voss (x.com/seldo) · Fuad Ali (x.com/tofuadmiral) · Paige Bailey (x.com/DynamicWebPaige) · PewDiePie · Jeffrey Hinton · Kent C. Dodds · Clay Christensen · Don Norman · Jack Ryan · Michael Grinich · Robert C. Martin · Grady Booch · Dax Raad · Matt Brockman · Tejas Kumar · Richard Feynman · Willem Pienaar · Nick Nisi (@nicknisi) · Zack Proser (@zackproser) · Swix · Jess Wang (@daRubberDuckiee) · Ankur · Dalton Caldwell · Michael Seibel · Paul Graham · Justin Kan · Emmett Shear · Karthik Ranganathan (x.com/karthikr) · Heather Downing (x.com/quorralyne) · Rob Dial (coachwithrob.com) · Cal Newport · Joseph · Dr. Beth Shapiro · Andrew Huberman · Svante Paabo · Ed Green · Sergey Zimov · Ellen Carlson · Olivia Moore (x.com/omooretweets) · Josh Elman (x.com/joshelman) · Eugenia Kuyda · David Pollon · Sarah Wang · Dr. Koa · Lukas Biewald
ToolsDeasy Labs · Collibra · SharePoint · OpenTelemetry · OpenInference · Claude · Arize Voice Pro · Gemma 4 · AlphaFold · WebAssembly · Ajax · Odysius · Qwen 3.5 · Heretic · E2B · WorkOS · GPT-4o Mini · Cleric · React · Fleet · Whisper · Handy · Codeex · Braintrust · Meko · Yugabyte · Git worktrees · Colossal Biosciences · OpenAI · Canva · Figma · Claude Sonnet 5.5 · Claude Opus 5.5
PapersDistilling the Knowledge in a Neural Network · Kano Model · The Mom Test · Care Reports · AVBD physics work