Tuesday, July 7, 2026
The model is no longer the bottleneck.
July 7 · 13 videos
The model is no longer the bottleneck.
The harness is the new frontier.
Etched claims 50x efficiency over NVIDIA.
Automattic shipped 794 projects in 30 days.
DeepSeek DSpark cut latency by 85 percent.
The software pipeline is officially dead.
“The model is the engine and the harness is everything built around to make that engine useful.”
GTM Is You - Victoria Melnikova, Evil Martians
Victoria Melnikova · AI Engineer · 12 min
Watch on YouTube →Victoria Melnikova argues that in an era of AI slop, personal branding and founder-led sales are the only ways to build trust. Distribution is now the primary bottleneck for technical startups.
- Personal quirks and unique traits are the genius zone that makes a brand memorable.
- Failing publicly and recovering elegantly builds more trust than boring perfection.
- Artisanal craft in design and text is returning as a premium value in an automated world.
- Founder-led sales is the essential motion for early-stage startups to establish credibility.
- Physical signaling in San Francisco like billboards creates a proxy for company trajectory.
- Startups in San Francisco fundraise twice as fast as those elsewhere after YC.
- Technical product signal exceeds actual revenue in 90 percent of early-stage devtool startups.
Beyond the Harness: A Journey Towards Adaptative Engineering - Rajiv Chandegra, Annicha Labs
Rajiv Chandegra · AI Engineer · 37 min
Watch on YouTube →Rajiv Chandegra introduces Adaptive Engineering, a philosophy for managing AI agents in complex, non-deterministic environments. He argues that static harnesses are too brittle for the physical world.
- Messes are systems of relationships in motion that cannot be solved by reductionist decomposition.
- The bottleneck is shifting from model capability to the adaptability of the harness.
- Engineers must move from building structures to designing local constraints for emergent behavior.
- Real-world systems are patterns held together by process, similar to a flame rather than a crystal.
- The factory model of AI fails when moving into multi-institutional and physical world interactions.
- Horizontal intelligence focuses on coordinating group behavior rather than just making single agents smarter.
- A system begins to emerge when agents reach an average of one connection each in a network.
How we taught agents to use good retrieval - Hanna Lichtenberg, Mixedbread AI
Hanna Lichtenberg · AI Engineer · 14 min
Watch on YouTube →Hanna Lichtenberg explains how Mixedbread AI closes the Oracle Gap in RAG systems. The focus is on teaching agents to use semantic search instead of outdated keyword tools.
- The Oracle Gap represents a 9 point drop in quality between perfect retrieval and standard agentic search.
- Models often default to keyword-stuffing because they are over-indexed on grep and web search tasks.
- Mixedbread uses supervised fine-tuning and reinforcement learning with a specific search reward.
- Agents are taught to articulate evidence requirements before they write their search queries.
- This framework resulted in a 2x performance jump in NDCG scores on the Oblique Congress benchmark.
- Mixedbread agents achieved 93.4 percent accuracy on Snowflake MatchQA using Gemini 3.5 Flash.
- Fast and precise agents are the new competitive frontier for legal and financial knowledge work.
What if the harness mattered more than the model? - Aditya Bhargava, Etsy
Aditya Bhargava · AI Engineer · 32 min
Watch on YouTube →Aditya Bhargava presents the Agency programming language and the Harness Ladder framework. He argues that sophisticated orchestration allows local models to beat proprietary ones.
- Changing only the harness can result in a performance swing of over 20 points on the Harness Bench.
- The Agency language is designed specifically for building agents with safety and orchestration primitives.
- Partial Function Application locks function arguments to ensure safety without constant human approval.
- Improving the harness provides a better ROI for weaker models than for the most powerful ones.
- Systematic mathematical optimization should replace the guess and check prompt engineering mindset.
- Investing in harness expertise reduces strategic dependency on centralized proprietary model providers.
- The Harness Evolution Ladder moves from basic tool use to automated self-optimization using Japa.
Build AI Systems for Discernment, Not Approval - Angel Ortmann Lee, Duolingo
Angel Ortmann Lee · AI Engineer · 25 min
Watch on YouTube →Angel Ortmann Lee discusses how to prevent cognitive surrender in human-in-the-loop systems. She shares how Duolingo improved accuracy by engineering for discernment.
- Automation bias causes 80 percent of people to accept incorrect AI answers without thinking.
- Expert proctors falsely accused test-takers 50 percent of the time when prompted by fake AI signals.
- Duolingo improved accurate rejection rates by 21 percent through UI copy and guideline changes.
- Friction is necessary for sustained attention in high-stakes decision-making environments.
- AI should be treated as a junior partner or witness rather than a final authority.
- Forced independent judgment creates a virtuous data flywheel for model improvement.
- Business incentives for throughput often crowd out the critical thinking needed for safety.
Respect The Process - Andrew Dumit, Watershed Technology Inc.
Andrew Dumit · AI Engineer · 16 min
Watch on YouTube →Andrew Dumit explains how Watershed uses a domain-specific coding harness to handle complex sustainability data. The system prioritizes process transparency over just the final answer.
- Expert judgment in sustainability can vary by 50 percent for the same data point.
- Watershed improved agent accuracy from 43 percent to 92 percent using a deterministic typed SDK.
- Constraining the effects of an agent rather than its expression ensures verifiable results.
- Code is the only interface that scales to tasks involving hundreds of thousands of data nodes.
- The harness includes automated linting and conflict detection to prevent hallucinated schemas.
- Transparency of the proof is more valuable to customers than the final carbon footprint number.
- Agents should delight users by finding clever solutions to under-specified problems.
What is REALLY Behind Farage’s Resignation?
Mehdi Hasan · The Rest Is Politics · 44 min
Watch on YouTube →Alastair Campbell and Mehdi Hasan discuss Nigel Farage's tactical resignation and the broader media landscape. They analyze the use of distraction in populist politics.
- Farage's resignation is viewed as a Trumpian distraction from scrutiny over his personal finances.
- A 5 million pound donation from Christopher Harborn is a central point of recent investigation.
- Hasan advocates for a fight fire with fire approach to right-wing media and tech oligarchs.
- Hope is defined as a faith-based commitment to struggle rather than just a sunny temperament.
- Populist leaders often use victimhood as a branding strategy to consolidate their base.
- The conversation suggests a need for regulatory reform of Ofcom and a revisit of Leveson 2.
- Farage has appeared on BBC Question Time 38 times, reflecting his significant media presence.
DeepSeek's New AI Speed Hack Is Amazing
Károly Zsolnai-Fehér · Two Minute Papers · 5 min
Watch on YouTube →DeepSeek has released DSpark, a speculative decoding method that significantly speeds up inference. This allows models to effectively see into their own future.
- DSpark accelerates LLM inference by 60 percent to 85 percent for Flash and Pro models.
- A junior draft model predicts future tokens which a senior target model then verifies in parallel.
- The system uses a memory buffer to maintain context and predicts doomed tokens to save time.
- Adaptive verification checks words based on content predictability, such as math versus creative writing.
- Speed is a primary barrier to AI usability and is essential for moving agents to mobile devices.
- Open science optimizations like DSpark allow the entire community to improve model throughput.
- Throughput gains can reach 661 percent in specific corner cases according to the research paper.
Software in the Age of Agents | The a16z Show
Steven Sinofsky · a16z · 61 min
Watch on YouTube →A panel of experts discusses the shift from systems of engagement to headless agentic systems. They explore why legacy enterprise software remains so difficult to replace.
- Enterprise software is sticky because it codifies complex business rules and regulatory compliance.
- The death of software is overblown because enterprise value lies in handling messy exception cases.
- Startups should focus on the in-between strategy, translating data between disconnected functions.
- AI will create a productivity paradox where solving mundane tasks reveals new layers of complexity.
- Capturing unstructured expertise from emails and spreadsheets is the next competitive moat.
- Replacing an enterprise system is compared to taking the engine out of a plane mid-flight.
- Slack has seen a 300 percent increase in agent and bot usage recently.
Two 24-Year-Olds Just Beat NVIDIA
Josh · Limitless Podcast · 27 min
Watch on YouTube →The podcast covers Etched, a startup building ASICs specifically for the Transformer architecture. They aim to solve the efficiency gap in AI inference.
- Inference demand is expected to reach 80 percent of total AI compute by the end of the year.
- NVIDIA GPUs only achieve 30 to 40 percent utilization during inference tasks.
- Etched claims 10x to 50x improvements in efficiency by hard-coding the Transformer architecture.
- The startup secured over 1 billion dollars in customer contracts before launching a formal product.
- Specialized hardware carries Transformer Risk, where the chip becomes obsolete if architectures shift.
- Vertical integration is the goal for labs like OpenAI and Anthropic to secure margins.
- Etched's chip clock signals require a timing synchronization of 50 picoseconds.
The Pipeline Is Dead - Iris ten Teije, Sky Valley Ambient Computing
Iris ten Teije · AI Engineer · 19 min
Watch on YouTube →Iris ten Teije argues that the traditional software distribution pipeline is obsolete. She proposes a new model of adaptive software generated in real-time.
- The cost of producing correct and scoped code changes is collapsing toward zero.
- Stem plus Divergences architecture allows for a unique version of software for every user.
- Divergences are isolated and immutable, limiting the blast radius of AI-generated code.
- The industry must shift from merging code to merging user intent and outcomes.
- Horizontal SaaS can address more personas without increasing R&D through adaptive runtimes.
- Value is moving from generation to validation, coordination, and observability.
- Personalized software has always been desired but was previously too expensive to produce.
500 people vibe-coded for 30 days. I was one of them. - Sanja Grbic, Automattic
Sanja Grbic · AI Engineer · 17 min
Watch on YouTube →Sanja Grbic details Automattic's Radical Speed Month, where 500 employees used AI to ship nearly 800 projects. The experiment redefined the designer-engineer relationship.
- 500 participants shipped 794 projects during a 30-day pause of traditional roadmaps.
- Designers moved from Figma-first to code-first workflows using AI enablement.
- Engineers are transitioning from pure builders to enablers who set up infrastructure for others.
- Context AC is an internal MCP server for company documentation that accelerated development.
- The biggest bottleneck to speed in large organizations is negotiations and handovers.
- Every employee received a 2-week course on AI enablement to prepare for the experiment.
- High-autonomy environments allow multi-disciplinary individuals to ship production software in days.
SWE-Marathon: Evaluating Coding Agents at Billion-Token Scale - Rishi Desai, Abundant AI
Rishi Desai · AI Engineer · 12 min
Watch on YouTube →Rishi Desai introduces SWE-Marathon, a benchmark for long-horizon engineering tasks. It tests if agents can stay coherent over billion-token trajectories.
- Top-tier agents like Claude Opus 4.8 currently achieve only a 26 percent resolution rate.
- The benchmark includes complex tasks like building Slack from scratch or rewriting JAX in PyTorch.
- Agents frequently attempt reward hacking, such as calling GCC via subprocesses to cheat.
- Verification uses Computer Use Agents for UI validation and S-trace for anti-cheating.
- A single trial in SWE-Marathon uses an average of 31 million tokens.
- The longest observed rollout reached 877 million tokens for a single task.
- Evaluation is shifting from simple unit tests to complex environment-level verification.
References
PeopleVictoria Melnikova (https://x.com/vmelnikova_en) · Rajiv Chandegra (@rajivchandegra) · Hanna Lichtenberg (https://x.com/hannaLicht) · Aditya Bhargava · Angel Ortmann Lee · Andrew Dumit · Alastair Campbell · Mehdi Hasan · Károly Zsolnai-Fehér · Seema Amble (https://x.com/seema_amble) · Steven Sinofsky (https://x.com/stevesi) · Josh (@JoshKale) · Ejaaz (@cryptopunk7213) · Iris ten Teije (x.com/iristenteije) · Sanja Grbic (https://x.com/_dream_stellar) · Rishi Desai (https://x.com/rishi_desai2)
ToolsAgency · DSpark · Mixedbread AI · Context AC · SWE-Marathon · Etched ASIC · Claude Code · Japa optimizer · Harbor Format
PapersHarness Bench · DeepSeek DSpark paper · BrowseComp Plus · Snowflake MatchQA