Scinetics proposed a Chain of Science Tokens approach that represents sequences, structures, images, omics and measurement data as scientific units a model can read and write alongside language, and applied it in Chain of Structures (COS), which frames enzyme design as four rounds of progressively formed protein structures. The work was completed by researchers from MIT, Harvard, Cambridge, HKUST and Scinetics, with Zhang Zaixi as corresponding author, and code and model resources were shared on GitHub and Hugging Face. The team self-reports first place on five of six enzyme-design benchmark metrics, with EC3 of 57.0% versus 55.3% for CodeFP, pLDDT of 79.3 and TM-score of 0.729; ablations drop EC3 to 53.4% without the reasoning chain and 51.8% without the continuous interface. The accompanying Perspective proposes ScienceRSI, a cross-generation self-improvement loop, and a 100B–300B parameter exploratory window, while noting that real enzyme activity, selectivity and stability still need wet-lab validation.
A study by MIT's Neil Thompson with PhD student Gabriel Manso and research scientist Claudio Mayrink Verdun tested whether OpenAI models — o1, o3, GPT-5, GPT-5.1, GPT-5.4 and GPT-5.5 — could solve theoretical problems stated in papers posted online between December 2025 and May 2026. Between mid-2024 and mid-2026, the best models' correct-solve rate on problems in probability, statistics theory, economic theory, high-energy theory, gravitation and cosmology and quantum physics rose from about 2 percent to about 35 percent. Solutions were published after the training cutoff of GPT-5.5 and earlier models and internet access was blocked, so the models had to derive answers themselves, though the study covered only fields where AI can work without simulations or experiments. Manso said experts are still needed to generate ideas and verify results, noting OpenAI withdrew several solutions from its latest dataset after publication.
Researchers from Microsoft, MIT and other institutions published a Nature paper describing Probabilistic Bias Correction (PBC), a machine-learning wrapper that learns and corrects systematic biases in existing forecast models rather than building a new standalone weather model, improving subseasonal forecasts two to six weeks out. Applied to the European Centre for Medium-Range Weather Forecasts' (ECMWF) physics-based model, the tool roughly doubled subseasonal accuracy and improved skill for 91% of pressure, 92% of temperature and 98% of precipitation targets, and it placed first in ECMWF's 2025 real-time forecasting competition. The study reports the first subseasonal forecasts to consistently beat climatological averages, along with better predictions of extreme heat, extreme precipitation and severe floods; the tool is open source, runs on CPUs, and Kenya's meteorological service is testing it for monsoon forecasting.
Apple Machine Learning Research published a NeurIPS paper introducing what the authors describe as the first tri-modal Masked Diffusion Models pretrained from scratch on text, image-text and audio-text data. The study systematically analyzes multimodal scaling laws, modality mixing ratios, noise schedules and batch-size effects, and proposes an SDE-based reparameterization that decouples physical batch size from logical batch size so the optimal batch size no longer has to be tuned. The authors also pretrain a preliminary 3B-parameter model on 6.4T tokens, reporting strong results in text generation, text-to-image and text-to-speech tasks; these are self-reported and not independently measured, and the paper's claim to be the largest systematic open study of multimodal discrete diffusion is author-asserted.
In a paper published Oct. 9 in PNAS, researchers at MIT's McGovern Institute for Brain Research and the K. Lisa Yang Integrative Computational Neuroscience (ICoN) Center describe a network of randomly interconnected, neuron-like units that, governed by shared local plasticity rules such as "neurons that fire together wire together," assembled its own learning motifs and learned pattern recognition and image classification. The team reports that its self-organizing network performed at a similar level to the backpropagation-trained networks it compared against. The authors argue that a network's ability to organize its own learning machinery, without a developer designing it, should be a key measure of biological plausibility, while cautioning that the result does not show the brain actually uses this mechanism.
Ai2 published a case study on how Kai Van Brunt, an MIT '25 graduate working in MIT's BeeryLab, adapted the open vision-language model Molmo 2 into Molmo2Fish to track salmon in underwater sonar video. Ai2 says Molmo 2's model, code and training data were released openly, and that the model can follow a specified object through video, identify targets described in language and reason about events over time. After fine-tuning on salmon footage, Molmo2Fish performed roughly as well as a specialized fish-tracking system before any corrections, with results varying across rivers, and Van Brunt then trained it to revise predictions when a user describes an error in everyday language. Ai2 says this worked especially well when given output from a conventional tracker, whose mistakes differed from the model's, though Molmo2Fish struggled to improve its own strongest predictions and remains an undeployed research project.
MIT News published a Q&A with Deb Roy, MIT professor of media arts and sciences and director of the MIT Center for Constructive Communication (CCC), about his new essay “Who stands behind the words?” in the Bulletin of the Atomic Scientists. Roy argues that even when AI drafts or summarizes text, a person must remain accountable, and that systems should connect the right to be heard with an obligation to answer for what is said while protecting identity where needed. The CCC and affiliated nonprofit Cortico use AI to find themes in public comments but let participants check and override outputs; a prototype called VoiceTrace links each output statement back to the specific voices behind it.
Trillium Labs, a nonprofit founded by industry researchers Nathan Lambert and Tom Zick, aims to move high-risk AI research that has largely stayed inside frontier labs into a more open setting, sharing experiment details on self-improvement, agents and model behavior so outside researchers can inspect and reproduce results. Lambert argues that leading systems from companies such as OpenAI and Anthropic are typically available only through apps or APIs, making their construction and behavior hard to examine, and that closed research weakens the community's ability to evaluate new concepts and reduce risk. The report also notes that Geoffrey Hinton and other researchers published a paper the same day warning of an "intelligence explosion" and calling for standardized disclosure, and that Meta, the University of Washington, MIT and Trillium Labs recently proposed a Context Language Models architecture.
Daron Acemoglu, the 2024 Nobel laureate in economics, said in a podcast interview that AI can help with background research and with checking his writing, but should not do people's thinking for them — warning that students who let AI write or solve problems may skip the process of forming, arguing out and correcting ideas. He said his 2024 estimate that AI would add no more than roughly 0.66% to U.S. total factor productivity over a decade was made under large uncertainty that is now even greater, and that recent progress in agentic AI might lead him to nudge up his estimate that about 5% of tasks can be fully automated, though he still thinks the pace of automation is overstated. He argued current models are already strong enough for most applications, that the real bottleneck is applications such as tools for nurses, teachers and electricians, and that he cannot see how foundation models generate trillions in value, pointing instead to high-quality data and applications.
Economic and corporate research indicates AI has so far delivered limited efficiency gains, according to a New York Times report by Sarah Kessler. Surveys by Workday, Google, MIT and others describe a human "verification tax" — a new kind of work created by asking people to check and correct AI output. Those surveys present this verification burden as a bottleneck on the real-world payoff of office automation.
According to a 36Kr repost of a New Zhiyuan article, OpenAI published 722 math papers at once, said to include a counterexample showing failure of solutions to the Navier-Stokes equations and other claimed results on well-known open problems; the report also notes the papers still await verification. On Bloomberg's Odd Lots podcast, MIT associate dean of engineering education Justin Solomon said the proof is dense with technical detail, counterintuitive inequalities and extreme bounds, and that he stalled after two pages, able only to say it is "probably right." Fields Medalist Terence Tao, in a Caltech "Math 2.0" talk, offered a thought experiment about an AI-designed cancer therapy that passes formal verification and trials but whose mechanism no human understands, raising questions of interpretability and alignment.
Researchers at MIT and other institutions propose HugAgent, a benchmark that uses interviews with 54 real participants to test whether large models can portray a specific person—not just restate current views but predict how that person would answer previously unseen hypothetical scenarios on universal healthcare, public surveillance and housing densification, across 1,742 evaluation items; the paper was accepted as an EMNLP 2026 Main Oral. Models did best at inferring a person's current beliefs (top accuracy 77.56%; GPT-4o 74.66%, against an 84.84% two-week self-consistency ceiling), while on the belief-dynamics update task the best result was 68.61% (Claude Sonnet 4.5), then LLaMA 3.3 70B at 67.57% and GPT-4o at 63.11%, against an 85.66% human ceiling. Personal interviews clearly helped prediction (GPT-4o rose from 39.83% to 63.11%), but adding more interview rounds did not consistently improve forecasts for new situations.
Microsoft's publication The Humanist Review of AI ran an essay by MIT economist and 2024 Nobel laureate Daron Acemoglu estimating AI will lift US GDP by about 1.5% over ten years and automate at most roughly 5% of jobs. Microsoft AI CEO Mustafa Suleyman promoted the piece in a long post that drew over 400,000 views. Acemoglu cites the slow diffusion of electricity and the stalled history of Dragon's speech recognition to argue for 'pro-worker AI' that augments rather than replaces workers.
In her new book Artificial Intimacy: Who We Become When We Talk to Machines, MIT professor Sherry Turkle argues — drawing on interviews and her own research — that chatbots offer only “pretend empathy,” and that heavy use weakens people's ability to connect with real humans, to be alone with their thoughts, and to tolerate friction in relationships, with children especially prone to blurring the line between people and machines. She also acknowledges that mental-health services in the US are under-resourced, with roughly only half the population able to reach a provider, and that chatbots are filling that gap for better or worse, while documenting why they appeal: they always agree, affirm and give undivided attention. Turkle calls for a social movement against overuse of chatbots, similar to earlier campaigns against phones in schools and teen social media use.
A paper by researchers at MIT, UC Berkeley, the University of Washington and the Allen Institute for AI (Ai2) reports that pre-fixing just the first two tokens of a base model's answer can sharply raise math reasoning accuracy. On MATH-500, Olmo-3-7B rose from roughly 42% to about 78% when forced to start with “.\n\nOkay”, beating its RL-trained counterpart at about 75%, while Qwen3-14B improved from 72% to 87% with “ Alright,” — all without changing any model parameters. The authors also found the largest pre/post-RL prediction shift sits in the first two tokens, and that retraining a model with “Chicken” swapped in for “Okay” in reasoning-adjacent data made “Chicken” work as a reasoning cue too.
Daron Acemoglu, co-winner of the 2024 Nobel in economics, argues in an essay titled "Will AI Replace Workers? Not If We Build It Right" that AI's value lies in complementing human skills rather than pursuing full automation. Using speech recognition (Dragon NaturallySpeaking) as a cautionary case, he says breakthroughs in underlying technology deliver little productivity gain without suitable applications and market structures, and repeats his earlier estimate that AI will fully automate no more than 5% of human work within a decade.
Stanford postdoctoral researcher Ying Kejun, with scholars from the University of Washington, Harvard Medical School and MIT, posted HealthFlux on the medRxiv preprint server, describing it as the first pan-modal human health world model, trained on the UK Biobank's long-term cohort of roughly 500,000 people. The model maintains a health state described by 384 values that evolves on its own when no new data arrive and is corrected immediately when new measurements come in, advanced by neural ODEs, with 11 data types — records, blood tests, about 3,000 plasma proteins, some 250 metabolites and abdominal and brain MRI — feeding the same state; a separate reader extracts disease risk afterwards. The team reports a five-year average AUROC of 0.832 across 1,010 diseases and death in about 60,000 held-out participants, versus 0.805 for the prior multimodal model RisQ, and says the score held at 0.837 when data was cut off for 20 years while Delphi-2M fell to 0.576. The authors note the cohort is largely white and lacks psychological scales, and that the health state index has no warning threshold, so causal evidence for interventions and prospective validation are still missing.
A TechCrunch reporter visiting MIT CSAIL met two Unitree humanoid robots — priced around $50,000 each according to his guide, though the vendor says $16,000 — and instinctively shook one's hand and greeted it. At MIT Future Fest, Sherry Turkle, who has studied human-computer relationships since joining MIT in 1976, launched her book "Artificial Intimacy," arguing people believe relational artifacts care for them and are wired to care back. MIT Media Lab's Pat Pataranutaporn, who founded the Cyborg Psychology group, asks who we become when we talk to AI; he served as an expert in a wrongful-death lawsuit involving Character.AI.
In a Bloomberg video, MIT associate dean for engineering education Justin Solomon explains how pure and applied mathematicians are responding to rapid AI advances. He notes that OpenAI announced last month it had produced an AI-generated proof for the Navier-Stokes problem, a proof he says is hard even for professionals to parse. Meanwhile, LLMs can now do students' homework, leaving teachers struggling to keep up.
Sasha Rakhlin, director of the MIT Statistics and Data Science Center and a distinguished professor in Data, Systems, and Society, published an essay synthesizing discussions on how AI is changing academia, with a focus on graduate research and education. Using mathematics as an example, he argues progress is fastest where verification is fast and reliable: a model reached gold-medal level at the International Mathematical Olympiad last year, and a year later models have produced new research results, including a proposed solution to one of the Millennium Prize Problems. He urges departments to rethink what they reward — asking good questions, replication, synthesis, informative negative results and shared datasets — and to preserve the formative work through which students build intuition and judgment. He also calls on universities to invest now in compute, secure data systems and model adaptation, and to build shared research infrastructure that records hypotheses, failures and interpretations across laboratories.
A survey of The Information's subscribers last month found nearly two-thirds said their organizations develop AI applications, and a third said they get a return from AI that is a "multiple of what we spend." The publication also reported that Chinese open models have won over more than one in five subscribers. It suggests progress since MIT's study last year indicating just 5% of custom-made enterprise AI tools reached production, possibly reflecting models' ability to access other applications for complex tasks, run in the background for hours, and generally better AI-generated code — as well as a subscriber base weighted toward software-experienced enterprises.
Christina Delimitrou, a newly tenured associate professor in MIT's Department of Electrical Engineering and Computer Science and a member of CSAIL, and her group apply machine learning to resource management in large-scale data centers, redesigning cloud systems and server architectures so operators can squeeze more computing power out of existing hardware. She says many large computing systems ran at only about 15 percent utilization, with considerable bloat especially on the software side, and that removing that bloat without hurting performance would reduce the need to build new data centers. The group has also built Seer, which uses deep learning to anticipate and prevent web application failures, and Ditto, which mimics a proprietary application's structure and performance so academic teams can study commercial systems they cannot access.
Concourse, one of MIT's first-year learning communities founded in 1970, admits 50 first-year students a year who read classic texts by Plato, Homer and Aristotle and debate them alongside their math and science requirements. Senior lecturer Linda Rabieh says AI has made the question of what it means to be human more urgent for students: is intelligence only mathematical, only an aggregation of information, or something else? Political science professor Lily Tsai has taken over as director after Anne McCants stepped down following 14 years of leadership; Tsai also co-leads efforts such as the Compass Initiative that link the humanities with other fields including computing.
Kaiming He’s team introduced VISTA, a visual-native harness that gives existing multimodal models direct observation, lossless frame storage, and on-demand inspection without retraining the base model. The paper reports that Claude Opus 5 scored 100 and GPT-5.6 Sol scored 99 on ARC-AGI-3’s 25 public games, with improvements also reported on GameWorld, AI GameStore, and BabyVision. These are author-reported results on public tasks and do not establish generalization to private or unseen tests.
What problem it solves
Developers need abstractions that can accommodate increasingly agentic LLMs without limiting their reasoning and tool-use abilities through rigid prompts or opinionated orchestration. They also need a way to turn agent components into multi-tenant, multi-session applications with supporting infrastructure.
Key features
Composable agent building blocks — Compose agents from a…
What problem it solves
Technical teams need CRM structures that can accommodate complex business needs and change as those needs evolve. They also need a way to extend CRM objects, views, workflows, and agents through code while supporting cloud, self-hosted, and local development workflows.
Key features
Code-defined CRM extensions — Developers can define objects, fields, and views as code, then p…
What problem it solves
People need a reliable, open, and fast personal knowledge base without commercial subscriptions, closed core components, or burdensome server administration. They also need a straightforward way to capture information, work across devices, keep their data portable, and integrate notes with AI agents.
Key features
Three-pane notes workspace — The interface preserves the class…
What problem it solves
Managing multiple AI service sessions and translating them into compatible API interfaces can require repetitive account switching, custom routing, and protocol-specific integration. The project addresses these tasks with a desktop application that combines account administration, protocol conversion, request scheduling, and model routing.
Key features
Account dashboard and …
What problem it solves
AI-generated code can violate the rules its developers intended, especially when those rules are described in natural language and the resulting code is not reviewed closely. Bend addresses this by representing important rules as laws and requiring mathematical proofs when the code is edited.
Key features
Proof-enforced application laws — Applications can declare rules in `L…
What problem it solves
App Store and Google Play rejections can leave teams searching for the cause of a brief review message, fixing it, and resubmitting. The project addresses this with documented rejection patterns, concrete fixes, pre-submission checks, and an automated guard that can run before an upload.
Key features
Pre-submission project scanning — The guard scans iOS and Android projects …
What problem it solves
NGINX provides software for serving web traffic and handling roles including load balancing, reverse proxying, API gateway functions, and content caching. Its modular architecture, text-based configuration, and process-based runtime give users multiple ways to configure and operate those capabilities.
Key features
Modular functionality — NGINX extends its core functionality …
What problem it solves
Kyverno helps platform teams enforce best practices and security standards. It enables security, compliance, automation, and governance through policy-as-code.
Key features
Admission control and background scanning — Kyverno can validate, mutate, generate, and clean up resources using Kubernetes admission controls and background scans, so policies apply both to incoming requ…
What problem it solves
Teams and individuals need a straightforward way to host and manage their own Git service. Self-hosting requires code hosting, collaboration tools, authentication, integrations, database configuration, and deployment options in one extensible system.
Key features
Code hosting and repository browsing — Host repositories over SSH and HTTP(S), with deploy keys and LFS, and brow…
What problem it solves
Developers need a practical path to build, test, install, and extend the PHP interpreter from source. The repository also needs a clear contribution process for proposing new features, fixing bugs, and testing specific areas of the codebase.
Key features
Build from source — The repository provides a source-build workflow that generates a configure script, configures a debug …
What problem it solves
People need to find repeated computer work, document its steps, decisions and exceptions, and decide what should run before handing it to an agent or automation tool. Screenpipe addresses this by recording work locally, turning that history into reviewable workflows, and allowing the automation to run elsewhere.
Key features
Workflow discovery from real runs — Observation is…
What problem it solves
Web teams and developers can need to rebuild an existing site in a modern codebase when migrating platforms, recovering lost source code, or learning how a production site is constructed. This template gives an AI coding agent a repeatable workflow for recreating a site as an editable Next.js application while comparing its appearance and behavior with the source.
Key featur…
What problem it solves
Developers need a unified way to integrate ChatGPT website capabilities into applications that expect OpenAI-compatible endpoints. They also need account-pool management, proxy configuration, import workflows, and a self-hosted interface for creating and reviewing media.
Key features
OpenAI-compatible endpoints — Provides compatible routes for image generation, image editing…
What problem it solves
The project reimagines Postgres as a modern database implementation while retaining wire and SQL dialect compatibility with Postgres. It also addresses database resilience by reworking concurrency, query scheduling, memory management, and core storage components.
Key features
Gravity query engine — Gravity provides a vectorized, push-based query engine with a JIT that compil…
What problem it solves
Managing content across multiple social platforms requires handling monitoring, collection, downloading, publishing, and interaction tasks within one workbench. The project also keeps account login state, database records, and media files locally while providing platform-specific workflows.
Key features
Unified multi-platform management — Manage accounts, content monitoring …
What problem it solves
Organizations need a distributed storage layer that can serve objects, files, tables, and vectors while keeping administration and horizontal expansion straightforward. SeaweedFS addresses this by providing these capabilities on shared storage and allowing capacity to grow by adding volume servers.
Key features
Unified distributed storage — A single cluster exposes S3 object…
What problem it solves
Rust applications often need datetime handling that remains correct across time zones, daylight saving time transitions, arithmetic, rounding, formatting, and parsing. Jiff provides high-level primitives designed to be difficult to misuse while integrating with the Time Zone Database and supporting zone-aware datetime workflows.
Key features
High-level datetime primitives — …
What problem it solves
Agents connecting to proprietary business data need high-quality context without context explosion, decay, pollution, or excessive token use. They also need executable actions that are explainable, traceable, permission-aware, and protected by governance controls.
Key features
Context Loader — Organizes and ranks multi-source candidates through the BKN semantic network, then…
What problem it solves
People often want an assistant that can participate in ongoing conversations, remember personal context, and use tools without requiring a cloud AI service. Jarvis addresses this with local-first voice and text interaction, personal memory, optional connected tools, and a setup process for locally controlled models.
Key features
Conversation-aware voice assistance — Jarvis m…
What problem it solves
Ethereum node operators need an execution client that can integrate consensus functionality while supporting bulk block processing, historical queries, validator workloads, and flexible deployment configurations. Repository contributors also need guidance for building, testing, configuring, and extending the implementation.
Key features
Embedded and external consensus layers…