# 2027: Who Owns the Next Experiment? > Published on ADIN (https://adin.chat/s/2027-who-owns-the-next-experiment) > Type: Article > Date: 2026-11-08 > Description: The most consequential prediction for 2027 may not be when AI becomes smarter than a person. It may be who gets to decide what happens next. A coding agent can propose another experiment. A drug-design model can propose another molecule. A brain implant can translate another intended movement. A... The most consequential prediction for 2027 may not be when AI becomes smarter than a person. It may be who gets to decide what happens next. A coding agent can propose another experiment. A drug-design model can propose another molecule. A brain implant can translate another intended movement. A security agent can propose another patch. In each case, the valuable step is moving from a plausible output to an action whose result teaches us something. Whoever controls that step can accumulate better evidence, improve the product and earn the right to handle more important work. **Our investment thesis is that falling inference costs will shift value toward companies that own consequential experiments and their results, rather than companies that merely generate more suggestions.** That is not a claim that models become commodities or that applications automatically win. A frontier lab could own the entire loop. A pharmaceutical company could keep its data inside. A hospital could refuse to delegate the decision. Those are the competitive questions. The forecasts below disagree most usefully about *how the next improvement happens*: more capable models, better experiments, a different architecture, or more control over what machines may do. Four bets turn those disagreements into a 2027 sourcing agenda. The dated hypotheses are proposed here for discussion, not predictions attributed to the people cited. \## Bet one: The research loop matters more than the next model launch Read the original [**AI 2027**](https://ai-2027.com/), published April 3, 2025. Its racing scenario puts a superhuman coder in March 2027, a superhuman AI researcher in August and artificial superintelligence in December. It uses fictional companies and offers racing and slowdown endings. These are concrete scenarios, not established future events. The mechanism is more interesting than the calendar. AI helps researchers build better AI; the improved system does more of the next research cycle; progress compounds. That is recursive self-improvement, or RSI. It need not involve a model spontaneously rewriting itself. It does require repeated improvements that increase the system's ability to produce further improvements. Dario Amodei makes the fast-loop case. At the [January 20, 2026 World Economic Forum discussion](https://www.weforum.org/meetings/world-economic-forum-annual-meeting-2026/sessions/the-day-after-agi/), he described coding and research accelerating the next model generation. The [published transcript](https://www.weforum.org/podcasts/radio-davos/episodes/ai-agi-dario-amodei-demis-hassabis/) records his estimate that models might be six to twelve months away from doing most, perhaps all, software engineers' work end to end. That clock began in January, not this November. Demis Hassabis challenges the inference from engineering progress to general scientific ability. He emphasizes originating the right question and experimentally testing an answer. [Fortune's January 23 account](https://fortune.com/2026/01/23/deepmind-demis-hassabis-anthropic-dario-amodei-yann-lecun-ai-davos/) describes his roughly 50% chance of AGI within the decade. The same reporting records Yann LeCun's sharper architectural objection: language models alone are insufficient; systems need representations that let them predict consequences and plan in the world. A longer horizon is not, by itself, a contradiction of Amodei. The substantive disagreement is whether scaling today's research loop closes the remaining capability gaps or whether important ingredients are still missing. Each researcher also has a commercial stake in a different approach. There is fresh evidence for the loop, but not yet for its strongest version. Anthropic's [release history](https://www.anthropic.com/news/models) lists Opus 5.5 on September 22, Sonnet 5.5 on September 28 and Haiku 5.5 on October 7. Its [Opus announcement](https://www.anthropic.com/claude-opus-5-5) reports stronger coding and 40% lower typical workload cost than Opus 5. These are vendor results, not an economy-wide productivity measurement. More importantly, [METR's November 5 assessment](https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assessment/) estimates a 1.9-times acceleration in Anthropic's internal-model benchmark trend after October 2025, with a 95% interval of 1.5 to 2.8 times. It judges AI likely explains most current acceleration. Yet it also says fully autonomous research contributions are not a large contributor to overall progress. Anthropic redacted much of the evidence METR considered most informative. This is why "ahead of AI 2027" is too broad. The scenario authors' [February retrospective](https://blog.aifutures.org/p/grading-ai-2027s-2025-predictions) put quantitative progress at about 65% of their scenario's pace; a labeled July update raised the preliminary estimate to about 75%. Their live scenario contains later forecast updates and a corrected graph. A [third-party tracker](https://ai2027tracker.com/) now places some internal Anthropic models near or ahead of one coding-horizon curve, but its post-April points are inferred from capability indexes, not new direct horizon measurements. Neither comparison grades the whole scenario in November. The denominators matter too. METR compares benchmark trends with historical trends. AI 2027's early-2026 research claim compares algorithmic progress with a hypothetical world without AI assistance. These are not interchangeable multipliers. **Proposed hypothesis for June 30, 2027:** human-directed research teams will remain essential to the strongest publicly substantiated examples of AI improving AI. The decisive counterevidence would be an independently evaluated, repeated cycle in which a system chooses a consequential research direction, runs the work and incorporates a validated improvement with no substantive human scientific decisions. An isolated coding task or a selected discovery would not settle it. If the hypothesis holds, experiment selection, reproducibility, evaluation and compute allocation are valuable products. If it fails, autonomous research companies deserve much greater attention, but their need for trustworthy evaluation and experiment capacity grows too. Compute, data, training time and chip production do not accelerate simply because code does. **The sourcing choice: would you fund another research agent, or the system that decides which of its experiments is worth running? What evidence would reverse your choice?** ## Bet two: Biotech will reward better decisions before it delivers faster cures Biology is the hardest test of the fast-loop thesis because an incorrect answer can remain plausible until a physical experiment disproves it. A molecule that binds its target may still fail on exposure, selectivity, toxicity, stability or manufacturing. A promising cellular result may not translate to a patient. That is not an argument against acceleration. It tells us where acceleration becomes valuable. Instead of generating a thousand additional candidates, a company might identify which ten experiments change a development decision, execute them reliably and learn from the failures. Anthropic's [October 27 drug-development update](https://www.anthropic.com/research/early-progress-drug-development) describes a revealing example with BigHat Biosciences. Models proposed antibody-fragment variants; the laboratory made and tested them. Along the way, the team questioned whether its stability and aggregation metrics were misleading, revised the measurements and tested the new predictions. The important event was not just improved molecular design. It was correcting the definition of success. That is an early, partner-reported case, not independent evidence of better clinical success. [Recursion's description of its platform](https://www.recursion.com/platform) makes a similar strategic argument: physical experiments generate proprietary data, models guide the next experiments and results return to the learning system. The open question is whether that loop produces transferable biological insight or becomes exceptionally good at its own assays. Batch effects, missing metadata, inconsistent protocols and differences between laboratories can turn apparent learning into measurement error. A defensible data business must know not only the result but how the sample, instrument and analysis produced it. Negative results matter because excluding them can teach a model to repeat expensive failures. Clinical development introduces a second loop with a different clock. The same Anthropic update says Bristol Myers Squibb plans to expand its Claude trial-planning collaboration in 2027. That is a dated plan, not a demonstrated increase in trial success. It also reports promising mouse results from Mithril Medicine and conditionally targets readiness to start human trials as soon as late 2028. A better 2026 model does not turn those programs into approved medicines in 2027. There is still substantial room to improve the journey. Better eligibility criteria can make a trial recruitable. Better site selection can find patients sooner. Cleaner endpoint data can reduce ambiguity. Manufacturing-record review and regulatory preparation can remove avoidable delay. But these improvements differ from shortening the observation period needed to establish durability, safety or clinical benefit. The [January 2026 EMA-FDA principles](https://www.ema.europa.eu/en/news/ema-fda-set-common-principles-ai-medicine-development-0) support AI use across the medicine lifecycle within a risk- and evidence-based approach. They do not authorize replacing a trial with a plausible simulation. Synthetic controls, surrogate endpoints and model-assisted analyses need justification for the particular decision, population and setting. This is also where the forecasters' disagreement becomes commercially concrete. Hassabis's emphasis on experiments suggests software alone cannot compress every step. Amodei's feedback-loop argument leaves open a much larger upside if models choose better experiments, not merely conduct the old ones faster. Anthropic's [November 5 scientific-usage study](https://www.anthropic.com/research/how-researchers-use-claude) finds use concentrated in analysis and writing, with physical data gathering less represented. Its modeled acceleration scenarios are illustrative, not observed gains or deadlines for 2027. **Proposed hypothesis for December 31, 2027:** the most persuasive near-term AI-biotech businesses will demonstrate faster or cheaper *development decisions* before they can demonstrate an AI-caused improvement in clinical success rates. Test the claim using prospective programs: cost and time to a predefined go/no-go decision, blinded or independent replication, and performance on experiments the model has not seen. The countercase is credible comparative clinical evidence showing that AI changes which therapies succeed, rather than only the speed of the process. For scouts, this creates two legitimate strategies. Own therapeutic assets and accept the long clinical clock in exchange for large upside. Or sell assays, experimental data systems, trial feasibility, recruitment operations and regulatory evidence tools to many asset owners. Those are different capital structures, not simply cautious and ambitious versions of the same company. Healthcare workflows offer a nearer revenue test. Documentation, referral coordination and evidence retrieval can save work without claiming to diagnose or treat. But time saved must include review and correction; a draft that creates more verification work is not automation. Patient-data access, liability, reimbursement and integration remain part of the product. **The sourcing choice: would you rather own a promising molecule, the assay that selects it, or the infrastructure that makes its trial executable? Which has the best evidence of pricing power by the end of 2027?** ## Bet three: BCI's first scalable product may be a clinical service, not a chip A brain-computer interface translates measured neural activity into commands or other outputs; some systems also stimulate tissue. Invasive implants, surface interfaces and non-invasive sensors make different tradeoffs between signal quality, procedure burden and durability. They should not be treated as one consumer-electronics category. The relevant forecasts do not all say "mass adoption next year." [Reuters' March 8 reporting](https://www.japantimes.co.jp/business/2026/03/08/tech/china-brain-computer-tech/) describes China's policy target of major BCI breakthroughs by 2027 and attributes a 2027 market estimate to CCID Consulting. In that same account, Yao Dezhong, director of the Sichuan Institute of Brain Science, places practical public use three to five years away. A policy milestone, a market estimate and an expert adoption horizon measure different things. Their tension is between technical progress and delivery at scale, not a direct disagreement about one deadline. A useful implant must do more than work during a demonstration. Someone must select patients, perform the procedure, calibrate the system, adapt it as signals change, provide rehabilitation, maintain software and support the patient when something fails. The economic unit may be a supported course of care rather than a device sale. That creates a less obvious connection to biotech. In both sectors, the company that owns the longitudinal outcome can learn something a model developer cannot infer from a static dataset. For BCI, that outcome is sustained function in daily life, not a brief peak in decoding performance. Clinical partnerships can become a source of evidence and distribution, but also an expensive service obligation. **Proposed hypothesis for December 31, 2027:** the clearest evidence of BCI commercialization will come from repeatable clinical delivery and sustained use, not from electrode counts alone. Assess new scale claims against longitudinal function, calibration burden, adverse events, retention, clinical capacity and a named payer. The hypothesis loses force if a broadly accessible product demonstrates durable unsupervised use with little specialist support and a reproducible payment model. The upside is not confined to the winning implant. Signal-quality monitoring, rehabilitation, assistive-device integration and clinical operations could serve multiple manufacturers. The countercase is that each device remains so specialized that suppliers cannot build a common product. **The sourcing choice: if an implant works but every additional patient requires substantial specialist time, should we fund the next implant or the company that makes the existing one deliverable?** ## Bet four: Security spending will follow delegated authority, not the date of a spectacular hack The more useful these systems become, the more access they need. A lab agent may reach proprietary experiments. A clinical assistant may reach patient records. A BCI service may update software attached to a person's daily function. An enterprise agent may deploy code. Capability becomes economically consequential when someone grants it permission to act. AI 2027 dramatizes that exposure with the theft of a fictional lab's model weights in February 2027. It is a scenario event, not evidence of a known upcoming attack. No independent source reviewed here establishes that a particular major hack will occur in 2027. There are real, narrower forecasts. [Gartner's February 2025 prediction](https://www.gartner.com/en/newsroom/press-releases/2025-02-17-gartner-predicts-forty-percent-of-ai-data-breaches-will-arise-from-cross-border-genai-misuse-by-2027) says more than 40% of **AI-related** data breaches will arise from improper cross-border generative-AI use by 2027. It does not say 40% of all breaches. Its [August 2026 forecast](https://www.gartner.com/en/newsroom/press-releases/2026-08-26-gartner-forecasts-the-market-for-securing-ai-will-reach-almost-5-billion-in-2027) projects almost $4.8 billion in spending on securing AI in 2027. That is a commercial research firm's market estimate, not realized spending or a promise that startups capture it. The defense case is not merely "AI makes attackers stronger." AI can find vulnerabilities and scale deception, but it can also review code and prepare repairs. A system with narrowly scoped access and a reversible action can be safer than a weaker one holding unrestricted credentials. Conversely, an excellent detector does little if it notices an irreversible action too late. For applicable medical cyber devices, [FDA's cybersecurity resources](https://www.fda.gov/medical-devices/digital-health-center-excellence/cybersecurity-medical-devices-frequently-asked-questions-faqs) already describe obligations around vulnerability management, updates and software bills of materials. Device classification and attack surfaces differ; these are not universal claims about every BCI product. They do show why safe maintenance must be designed into the service, not added after commercialization. Politics can change who buys these controls. [Reuters' post-midterm reporting, updated November 5](https://www.reuters.com/5f374b32b07b/legal/government/democrats-seize-congress-sweeping-midterm-repudiation-trump-2026-11-04/), reports Democratic wins for control of both chambers and the prospect of deadlock with President Trump. The incoming majorities take office in January. That establishes a different oversight and funding environment, not an enacted AI law or a known biomedical budget. Congressional scrutiny could create demand for evidence and control even if comprehensive legislation stalls; that is our hypothesis, not a reported election result. Gartner's [October 20 technology agenda](https://www.gartner.com/en/newsroom/press-releases/2026-10-20-gartner-identifies-the-top-strategic-technology-trends-for-2027) likewise emphasizes agent identity, context infrastructure and coding controls alongside speech reasoning and world models. It identifies areas to explore in 2027. Its numerical agent-breach forecast is for 2028, and its sovereignty forecast for 2030. Neither should be pulled forward to fit next year's narrative. **Proposed hypothesis for December 31, 2027:** in consequential deployments, security tools that can demonstrate reduced blast radius or recoverable execution will be easier to justify than tools whose main evidence is more alerts. Test this in actual procurement: does the buyer pay for constrained permissions, safe release, data-boundary enforcement or recovery, and does the control survive a realistic adversarial exercise? The countercase is that existing platforms bundle these features effectively enough to eliminate a standalone startup market. This points toward agent identity, sensitive-data controls, secure device operations and recovery infrastructure. It does not imply that every new security layer is a company. Buyers may prefer one accountable incumbent to several tools that each observe only part of the workflow. **The sourcing choice: would you fund a tool that identifies more attacks, or one that lets a customer safely delegate more valuable work? Can the founder demonstrate the difference?** ## What would change our portfolio? These bets are linked, but they do not require the same answer. AI research could accelerate dramatically while clinical validation remains slow. BCI could deliver profound benefits to a narrow population without becoming a mass market. Defenders could gain from AI while badly governed agents create new losses. The central wager is about ownership of learning. Does a company control a recurring decision, the experiment that tests it and the result that improves the next decision? Or does it rent a model, produce a suggestion and hand the valuable learning to someone else? For the call, choose the bet you most disagree with. Bring one company, the paying customer, the result it must demonstrate by 2027 and the evidence that would make you pass. The aim is not agreement about the year intelligence arrives. It is a better reason to know which founder to call next. *Forecasts and evidence reviewed November 8, 2026 (ET). The four dated hypotheses are new editorial propositions for debate. They are not a consensus forecast or predictions attributed to the cited sources.*