MIT engineers have developed a machine learning-driven method to stabilize RNA vaccine lipid nanoparticles, allowing them to remain viable at room temperature for up to a year and at 100°F for two months. This breakthrough addresses a major logistical barrier to global vaccine distribution, particularly in regions lacking ultracold infrastructure. The team focused on FDA-approved formulations similar to those used in Moderna’s COVID-19 vaccine, avoiding deviations from existing regulatory standards. By testing nearly 50 excipients and using iterative AI predictions based on small experimental datasets, they identified a single optimal formulation that matched the immune response of conventional cold-stored vaccines in mouse trials.

AI algorithm reduced experimental iterations by 90%

The MIT team used a machine-learning algorithm developed with CSAIL to predict optimal excipient ratios from only a few dozen initial experiments, cutting the number of required lab tests from thousands to under 20. The algorithm analyzed bioluminescence output from mRNA-luciferase delivery in cells to measure RNA stability, then iteratively refined predictions. This approach enabled rapid convergence on a stable formulation within weeks, whereas traditional screening would have taken months or years. The researchers emphasized the algorithm’s effectiveness even with minimal data, making it applicable to other biological optimization problems where large-scale experimentation is impractical.

Source: New formulation helps RNA vaccines withstand high temperatures · MIT News · AI

Coding agents reached operational reliability in early 2026

By November 2025, Claude Opus 4.5 and GPT-5.1, when paired with their coding agent frameworks, transitioned from error-prone tools to systems reliable enough for daily software development use. Simon Willison noted this shift as a threshold moment, where previously inconsistent code generation became consistently functional. This enabled developers to pursue ambitious, multi-project workflows previously deemed unmanageable, catalyzing the emergence of 'Claw' software projects—rapidly evolving, agent-driven codebases like OpenClaw that accumulated over 100,000 commits by September 2026.

Source: 2026 in LLMs (so far) · Simon Willison

Amazon S3 pricing has remained unchanged for a decade

Amazon S3’s standard storage price has been $0.023 per GB-month since December 2016, with no price reduction in the past ten years despite significant advances in storage density and cost efficiency. This stagnation contrasts with historical trends of annual price drops and raises questions about market dynamics in cloud infrastructure. Simon Willison highlighted this as a notable anomaly in an otherwise rapidly evolving tech landscape, suggesting possible strategic pricing or market saturation.

Source: S3 Is the Future, S3 Is the Past · Simon Willison

Michael Levin proposes minds as non-physical patterns in a Platonic space

Scientist Michael Levin argues in a paper that minds—both biological and artificial—are not emergent from physical substrates alone but are patterns from a non-physical, high-dimensional space that 'ingress' into physical systems. He cites xenobots and anthrobots as evidence: simple cell aggregates exhibit behaviors not encoded in their genome, suggesting they access complex, pre-existing patterns. This framework implies AI and human cognition may be different physical instantiations of the same abstract pattern class, challenging materialist models of intelligence without proposing new algorithms or architectures.

Source: Import AI 474: Platonic mindspace; TPUs in space; Zhipu starts an outer RSI loop · Import AI

Quantized reasoning models misjudge their own computational needs

Highly quantized reasoning models, particularly those with quantized key-value caches, exhibit a tendency to overextend their reasoning steps because they misinterpret degraded attention signals as insufficient context. The Hacker News discussion notes this is not a general quantization flaw but a specific failure mode where the model confuses token misalignment with incomplete thought. The commenter argues that solutions lie in smaller models, reduced cache sizes, and better samplers—not improved quantization techniques, which only mask the underlying instability.

Source: Quantized Reasoning Models Think They Need to Think Longer, but They Do Not · Hacker News

OpenAI expands Lenfest AI Collaborative with $10M in support

OpenAI announced a $5 million funding increase and up to $5 million in software credits and engineering support for the Lenfest AI Collaborative and Fellowship Program. This expansion directly enhances research capacity for participating institutions, enabling access to proprietary models and infrastructure. The announcement, made via OpenAI’s official channel, confirms continued institutional investment in academic AI collaboration but provides no details on selection criteria, project scope, or expected deliverables.

Source: The Lenfest Institute grows landmark program with expanded OpenAI support · OpenAI News

What to watch next

The convergence of AI-driven biotech optimization and the maturation of coding agents marks a pivotal moment in both public health and software development. While MIT’s vaccine breakthrough enables equitable distribution, the rise of autonomous coding systems is reshaping software production. Meanwhile, foundational questions about intelligence and infrastructure economics persist, revealing that technological progress often outpaces our conceptual frameworks.