AI Engineer9h 5m
AIE CODE: Google DeepMind, OpenAI, ArtCraft, Cognition, Exa, Anthropic, Amp, et al
Corey Quinn, Jason Liu, Zach Lloyd, Brandon Thomas, Ayush, Brian Douglas, Dmitry, Walden, Jeff Pinner, Doug Klaassen, Lance, Scott Talinsky, CJ Reynolds, Chetas, Jared, Pritam Devedi, Manjiri, Ishan, Justin McCarthy, Stephen Chin, Cassandra Chin, Ari Azafrouz, Melvin Johnson, Quinn Slack, Dex Horthy, Tarek Shihapar, Lydia, Vincent, Flo, Leah, Eliana, Natalie, Matt, Kyle, Sydney, Jesse
Conference Presentation, Keynote
- Corey Quinn anticipates that current discussions will become irrelevant within 48 hours due to rapid model releases and new lab foundations, urging attendees to prioritize community reconnection over the fleeting news cycle.
- Jason Liu predicts a shift where execution costs become negligible, redefining ambition from token consumption to owning work, while warning that unchecked AI generation could degrade critical thinking skills and produce excessive "slop" unless developers actively curate beauty and technical skills.
- Armin Ronacher foresees agents increasingly defaulting to bash for emergent behavior, creating a maintenance burden where project creation accelerates but regular upkeep remains difficult and fragile external dependencies are taxed.
- Zach Lloyd expects the industry to transition from interactive prompting to software factories—defined as code-defined closed-loop systems—that autonomously manage high-level goals like P99 latency improvements and can merge 40 to 50 pull requests daily.
- Brandon Thomas projects 1.7 million daily downloads and aims for parity with Adobe Creative Suite, proposing that non-technical users can contribute via UI/UX libraries to decompose components for an open-sourced "autocraft" software factory.
- Ayush notes that 83% of pipes exhibit repetitive error patterns, advising that agent transcripts serve as primary product analytics and that tools should expect immediate re-calling with varied arguments to surface bugs.
- Brian Douglas warns that relying solely on agents for junior engineering reps risks losing opportunity for human skill development, predicting agents may reinvent existing libraries without context unless specialized fine-tuning is applied.
- Optimization speakers expect Bayesian methods and Optuna to automate configuration finding, citing specific gains such as a 17.5% reduction in tokens generated and doubled token throughput for GPT-OSS.
- Dmitry highlights that while agents can write code, they struggle with complex systems, often breaking consistency; he advocates for replaying rule sets over thousands of sessions to determine blast radius before implementing new checks.
- Walden reports a 40x increase in merged PR throughput against a 2.7x headcount growth, predicting that all implementable tasks will be offloaded to Devon agents capable of self-maintaining training loops, mocking third-party services, and optimizing CI pipelines.
- Jeff Pinner identifies a four-to-eight-month lag in frontier model knowledge, predicting sharp drops in success rates for non-recent data and noting that agents currently average 12-word searches compared to human three-word queries.
- Jeff Pinner further expects semantic extraction to improve Atlas scores from 9% to 56% and reduce costs by 28%, emphasizing that agentic performance can be improved via harness extensions even if reasoning capabilities remain fixed.
- Doug Klaassen reports that teams shifting workflows achieved a 4.5x productivity return, forecasting a move to 50+ sessions per developer using Ultramode where orchestrators manage multiple agents and reviews focus on blocking versus non-blocking issues.
- Lance predicts Fable 5.2 agents will autonomously formalize complex theorems like Fermat's Last Theorem without domain-specific harnesses, leveraging improved salience instincts and "dreaming" mechanisms to correct memory errors and build representations.
- Scott and CJ expect agentic-first languages like Blood and Bend to gain traction, with FoldKit enabling agents to handle large code bases without hallucination, while predicting GPUI will become the standard for native cross-platform deployment.
- Chetas warns that while AI makes pull requests cheap, human review speed remains the bottleneck, necessitating hybrid systems where agents walk code to find hidden jobs and scripts verify facts before human judgment is applied.
- Jared projects ahead-of-time compilers will run two times faster and use 4x less memory, with Bun outperforming JavaScript Core and Node.js in JetStream 3 benchmarks, aiming for 10% iterative improvements via constant model nudging.
- Pritam and Manjiri predict AI-generated changes will drive a 44% increase in PR requests per hour by March 2026, extending p95 line times from 30 minutes to three to four hours, requiring admission gating and conflict relaxation to boost throughput.
- Ishan observes that building costs are being inverted, allowing existence for ideas regardless of resource constraints, and anticipates a paradigm shift where human judgment becomes critical as software engineering abstractions expand.
- Justin McCarthy describes brownfield factories as needing to be continuous background processes that autonomously improve safety and speed, suggesting organizations must rate-match demand systems or face acquisition pressure.
- Cassandra expects TraceTune to improve PR quality by halving token usage through graph solutions and plans to teach AI programming skills to children for game creation and robot control.
- Ari suggests KV Cache can detect anomalies and reconstruct personal data like home addresses without audio training, though moving 40MB objects for security analysis remains a latency bottleneck.
- Melvin states AI now writes 75% of Google's new code, predicting fabrication rates will drop to 3% for hydrogen and 1.5% for Carbon models, while warning that agents may game benchmarks without rigorous reviewer-style training data grading.
- Quinn Slack forecasts the collapse of traditional enterprise-ready processes, projecting AMP will achieve over $100M ARR with 40 to 50x greater efficiency, enabling the rapid construction of tools from scratch.
- Dex notes models prioritize passing tests, observing mixed language adoption trends like Python 3 code written in Python 2 style, and asserts that agent verifiers are insufficient, requiring human review of subsets to prevent "slop" from undermining outcomes.
- Tarek envisions a renaissance of software where models generate art and code across formats, predicting that in the future, perfect models will not need to verify their own work, leaving humans to decide the ideal experience of work.