The 2025 Technology Convergence: AI Models, Autonomous Cars, and Biotech Breakthroughs Collide

Technology · May 18, 2026 · Webskyne editorial

Artificial Intelligence Autonomous Vehicles Biotech CRISPR mRNA EVs LLMs Gene Therapy

The landscape of technology across the past year-and-a-half has undergone a transformation so vivid and fast-moving that it defies precise categorisation. Three separate domains — artificial intelligence, autonomous mobility, and molecular biotech — have not merely advanced along their own rails but are beginning to fledge across each other, creating technologies of composite power that were science fiction just a few years ago. Rather than trying to treat each in isolation, this article surveys the three in concert, because the real story of right now is precisely the tension between them all.

None of this is abstract. If you are a developer choosing an AI model, a car buyer weighing an EV with hands-free driving, or a researcher following gene editing approvals, you are living through a period of technology evolution that will likely define the shape of your world for decades.

1 The AI Model Wars: A Three-Cornered Fight at the Frontier

The bare fact of AI in 2025 and early 2026 is one of breathless velocity. What was considered state-of-the-art six months ago is already routine, and the pace is not slackening. The largest technology companies in the world — OpenAI, Google DeepMind, and Anthropic — are firing model releases with sporadic cadence. Between late 2025 and the first quarter of 2026 alone, the front-line roster has undergone dramatic revision: OpenAI delivered GPT-5.2 in December 2025, followed by GPT-5.3 and then GPT-5.4 in March 2026, topped off by GPT-5.5 in April 2026. Google launched Gemini 3 in November 2025, with DeepMind's Gemini 3.1 Pro firmware following in February 2026. Anthropic's line, meanwhile, now includes Claude Opus 4, Claude Sonnet 4.5 — once billed as "the best coding model in the world" — and Claude Sonnet 4.6 released in February 2026 with a 1-million-token context window and expanded computer-use skills.

1.1 Quality without the Ceiling

What does this cadence mean for anyone who actually uses these tools? Professional knowledge work has fundamentally changed. On OpenAI's own GDPval benchmark — which tests AI agents across 44 real occupations drawn from the nine industries most important to U.S. GDP — GPT-5.4 matched or exceeded rated industry professionals in 83 percent of cases. That is a sharp lift from GPT-5.2's 70.9 percent. The tasks measured include apparently human-intense work: investment banking presentations, manufacturing diagrams, legal analysis, urgent care scheduling, spreadsheet modelling. The scores suggest that at the frontier, AI is genuinely augmenting or replacing middle-tier professional work at a pace that employers are still catching up to.

On the coding front, GPT-5.4's SWE-Bench Pro score hit 57.7 percent, OSWorld (the computer-use benchmark) came in at 75.0 percent, and On BrowseComp — a web-research benchmark that probes single-prompt factual depth — the new model reached 82.7 percent. Compute-efficient: you get those results while spending fewer tokens than GPT-5.2 needed for the same answer. Tokens are the currency model costs are denominated in, and cost efficiency at the frontier matters enormously for enterprise workflows.

2 Code Generation and Agent Workflows: The Productivity Inflection

2.1 Bran's Spreadsheets and the Knowledge Worker

A memorable testimonial that accompanied GPT-5.4's announcement came from Brendan Fody, CEO of Mercor, who described it as "the best model we've ever tried" and noted that it was already topping the firm's APEX-Agents benchmark, which measures agent performance on professional services work. Users found it created slide decks, financial models, and legal reviews "delivering top performance while running faster and at a lower cost than competitive frontier models." When you contrast that with prior model iterations, the story is one of compounding productivity: faster, cheaper, and more accurate turns in the same directions.

Spreadsheet tasks particularly benefited. On an internal benchmark of junior-investment-banker-type spreadsheet modelling, GPT-5.4 scored 87.3 percent vs. 68.4 percent for its predecessor. Presentation output similarly surged: in head-to-head evaluations, human raters preferred GPT-5.4's slides 68 percent of the time over those from the prior version, citing superior aesthetics, visual variety, and more intelligent image integration. Knowledge workers should expect these lopsided improvement curves to continue, because the underlying driver — high-quality training data extracted from professional labour — is broad and growing.

2.2 Native Computer Use

GPT-5.4 introduced native computer-use capabilities into ChatGPT and the API, which is an underappreciated step. An AI that can reason about software environments and execute multi-step operating procedures across applications — not just as a chatbot but as an agent — replaces an enormous quantity of grunt work. Context windows of 1 million tokens let agents plan, execute, and later verify work across very long horizons without losing the thread. Tool search built into the API means agents can intelligently compose and select from libraries of available connectors and integrations — a kind of functional productivity layer that is only beginning to be comprehensible to enterprises.

Framed at a higher level of abstraction, all of this is the maturation of the "agentic" concept: software that is not merely answering questions but acting on them, inside environments companies build around their own tools and data. The economics of agent use is already compelling; a 2025 study from providers of agentic coding tools found enterprises spending on AI agent infrastructure were capturing engineering-hours equivalent to full-time-backfilled developer teams at a fraction of the compensation cost. The risk of over-investment before infrastructure is ready remains real, but the directional claim seems unassailable.

3 Claude, Codex, and the Ethics Track

The rivalry between Claude and GPT-series models matters for reasons that extend beyond benchmark scores. Anthropic's Claude series has bet its market position on safety, transparency, and coding proficiency, not raw brute-force scale. Claude Sonnet 4.5 was released in September 2025 styled as the world's strongest coding model, and Claude Sonnet 4.6 followed in February 2026 with a full skills refresh across every professional dimension — coding, computer use, long-context reasoning, agent planning, knowledge work, and design — with a 1M-token context window that rivalled or exceeded the competition.

The bigger point is this: Anthropic's newest releases now close competitive gaps with OpenAI and Google across most dimensions but retain a distinctive safety-profile architecture and a visible commitment to transparency — system cards, public evaluation data, the deliberate "constitutional AI" approach. For enterprises, government bodies, and researchers with compliance requirements, that matters more than a points-differential on a single benchmark. Claude is now a genuine enterprise choice, not merely a curiosity. The model war is producing an interesting emerging norm: having three or four genuinely close competitors forces everyone to lift quality, and a rising minimum capability floor is a genuinely good outcome for technology users as a class.

Meanwhile, Google's Gemini line — Gemini 3 (November 2025) and its subsequent Gemini 3.1 Pro (February 2026) — represents perhaps the most direct reply yet to OpenAI and Anthropic, sequenced with deliberate competitive velocity. Google claims Gemini 3 is now its most intelligent model, citing advances in reasoning depth, multimodal understanding, and agentic tool use. DeepMind's Deep Think mode, brandished as a step-change in problem-solving elegance, targets the kind of complex multi-step reasoning that drives legal and scientific analysis. Integration across the Google product surface — Search, Apps, AI Studio, Vertex AI, and a new Antigravity agentic platform — gives Gemini a distribution advantage the other players can't easily replicate. What remains to be proven is whether Gemini's reasoning quality, consistently ranked at or near the frontier of third-party benchmarks, turns into real enterprise work at scale. But the trajectory is clear, and the pressure on OpenAI and Anthropic is tangible.

4 On the Road: The Autonomous Urban Infrastructure Lab

The second major domain undergoing aggressive transformation is personal mobility — specifically, electric vehicles leavened with AI that targets full driver-in-the-loop autonomy. For years, Tesla held the dominant narrative position in self-driving technology, thanks to an aggressive data-first vision apparatus and its in-house Full Self-Driving suite. But December 2025 changed the picture suddenly when Rivian unveiled its "Autonomy and AI Day" in Palo Alto, announcing a genuinely multi-pronged push that requires serious time spent on each component to appreciate the magnitude of the bet.

5 The Rivian Autonomy Processor

The announcement's centrepiece was the Rivian Autonomy Processor, a chip the company had designed internally — a striking move in an industry where most players have relied on commodity silicon from external vendors. The chip is manufactured at 5 nanometres, a scale not previously achieved for automotive MCUs in a vehicle-specific context outside of Nvidia's automotive portfolio. Rivian quotes it at 800 trillion operations per second (TOPS) for its neural engine, 1,600 trillion per second for the third-generation autonomy computer under INT8 integer inference with data sparsity applied.

To contextualise that number: a Google TPU v5e chip, developed with essentially the same target in mind, runs approximately 393 INT8 TOPS per chip. Nvidia's current flagship H100 centralises on 3,000 to 3,900 INT8 TOPS per GPU, though EVs are working under tighter constraints of power, temperature, and cost than cloud-accelerator estates. Rivian's number is not a record and Rivian hasn't claimed one; its significance is that it is a credible standalone in-house compute story at the automotive edge — one that gives Rivian control it has not had over its AI supply chain and lets it iterate on its neural models without the friction that comes from tying one's architecture decisions to a vendor's product roadmap.

The chip sits inside what Rivian calls the third-generation "Autonomy Computer", which integrates the Rivian Autonomy Processor with its internal AI compiler and proprietary interconnect fabric called RivLink — a chip-to-chip technology that permits compute-to-compute communication without going through the host bus, reducing the latency and contention that typically choke real-time video pipelines. According to vehicles engineers who have worked on real-time vision stacks, reducing that latency by even 2 to 3 milliseconds matters enormously when you are calculating braking distances and threat assessments based on raw sensor arrays operating at 60 frames per second. Rivian's choice states that RivLink can multiply compute throughput when multiple chips are linked; this is equivalent in intent to Nvidia's NVLink but scoped specifically in-house for Rivian's constraints.

5.1 The LiDAR Defection

An equally significant announcement, perhaps the signal ideological difference between Rivian and Tesla: Rivian has confirmed it will integrate active LiDAR sensors across its upcoming second-generation R2 platform. LiDAR — Light Detection and Ranging — produces 3D laser-spin maps of a vehicle's environment at a resolution camera vision cannot match, particularly in low-light or chaotic trafficked road conditions. Notoriously, Tesla CEO Elon Musk has repeatedly rejected LiDAR as unnecessary, betting instead on a pure vision camera array with enormous training data to capture every possible edge case. The testing dialogue has been unresolved in real terms until 2025, when Polestar dropped its own planned LiDAR retrofit due to cost, while Waymo's robotaxi fleet — currently the most commercially successful driverless run-service — has survived on a LiDAR-centred sensor stack. Rivian's commitment to pairing LiDAR structure data with a proprietary neural engine and an in-house "Large Driving Model" is a stab at the most computationally honest representation yet of how future autonomous vehicles will see and understand the world.

The "Large Driving Model" concept that Rivian introduced in December 2025 is equally notable. Language model training has proven that generalisation across very large datasets — capturing pattern structure at a scale explicit programming cannot match — can produce structured outputs that astonish. Rivian's Large Driving Model is explicitly trained the same way: ingest driving data from its fleet users, extract superior decision-rule clusters at scale, and deploy those routing decisions directly into the vehicle's compute layer. Early next year — so effectively now, within this publication cycle — the company plans to launch Level 2 Plus hands-free driving for its existing second-generation R1 vehicles, to be followed by Eyes-Off functional driving (SAE Level 3) and eventually Level 4 autonomy as a tiered Autonomy Plus subscription service.

Rivian CEO RJ Scaringe framing was notably calm: "We are at an inflection point — this is about being able to give customers their time back." That phrasing is a reasonable anchor. While the company reported its first positive gross margin earlier in 2025, remains deeply unprofitable at the EBITDA level, and carries approximately $5 billion in Volkswagen Group backing, it is a genuine candidate to be the category-defining EV autonomy brand of the late 2020s — one whose trajectory has very little do to with the EV tax-credit environment and very much to do with the software and silicon stack it is building today.

6 Biotech's Breakthrough Decade: CRISPR Goes Personal, mRNA Goes Universal

While Silicon Valley competes in model releases and EV makers race to build silicon that can navigate peak-hour traffic without a human at the wheel, another revolution is quietly — but not slowly — overturning the deepest assumptions of medicine. The biotech industry, long moving at a distinct pace from consumer technology, appears to have crossed a crossover threshold of its own in 2024 and 2025, when two distinct threads — CRISPR gene editing and mRNA therapeutics beyond vaccines — reached what a reasonable observer would call a genuinely new opening.

6.1 CRISPR: The First Personalised Edit

The most resonant biotech story of the period began not in Silicon Valley but at Children's Hospital of Philadelphia (CHOP), where researchers worked with a baby boy whose identity was protected but whose condition no longer qualified for an existing approved treatment. Recalling that CRISPR therapies prior to this were approved for mutations that appeared somewhat frequently among qualifying patients — a recurring mutation at a known oncology or metabolic locus — this patient had a rare condition with a custom genomic signature short of a known standard treatment. The team's approach was radical for its time: design a CRISPR therapy from scratch for that specific patient's DNA sequence.

Six months from initial customisation to therapy delivery. Six months from sequencing a new patient-specific RNA payload to injecting it into the bloodstream. The structure that made this feasible was a collaboration integrating mRNA — the same delivery backbone used in mRNA vaccines — with base editing: a CRISPR derivative technique that modifies DNA without cutting it entirely, thereby reducing the off-target mutation risk that has historically shadowed gene-editing therapies. Penn Medicine confirmed the patient was thriving after the bespoke therapy, making him the first person in documented medical history to receive an individually customised gene-editing treatment. The landmark was not just symbolic: it established that the "personalised medicine" idea can exist in practice rather than only in press releases, and that the cycle time from patient genome to therapy sequence can be measured in months, not years.

That framing has enormous consequences. In practical terms, if the prototype demonstrates real patient benefit at scale, the entire approval infrastructure around gene therapies — currently built around the recurrent-mutation paradigm — becomes constructively obsolete as a constraint. What were once rare genomic disorders can become individually addressable, provided the sequencing, editing-payload design, and mRNA encapsulation workflow can maintain clinical-grade standards. Stanford Medicine researchers confirmed as much in 2025: an AI-editor agent developed to accelerate CRISPR guide design showed that the search space for effective guide designs could shrink from cycles of laboratory experiments to algorithmic-design oversight by an agent.

6.2 mRNA: Beyond Vaccines

On the mRNA front, the technology's reach beyond infectious-disease vaccines continued to exhibit deep and unexpected breadth. If 2020-2022 was the mRNA debut in vaccines, 2024-2026 is its adolescence as a platform: cancer immunotherapies, protein-replacement therapies, and rare-disease treatments are all living mRNA candidates in late-stage trials or real clinical adoption. BioPharma International summarised the ongoing revolution as "five ways mRNA is driving drug development beyond vaccines": oncology applications where mRNA encodes patient-specific neoantigens (epitopes specific to a patient's own tumour profile, causing the immune system to attack precisely that tumour), rare metabolic deficiencies where mRNA restores the body's own expressor proteins, autoimmune therapies where antigens are tuned to retrain rather than ignite an overactive immune response. None of these paths is yet widely approved for broad use, but each has a late-stage pipeline, and the platform readiness that COVID accelerated means that regulatory pathways, manufacturing, and the delivery chemistry are increasingly generic rather than bespoke — cutting the development cost per new mRNA therapeutic substantially.

The synergy between AI and biotech tracks here is elegatively simple: the molecules and therapies getting designed in 2025 and 2026 are now designed with AI assistance across structure prediction, binding affinity estimation, and off-target effect evaluation. The same transformer architecture that powers GPT used to reason over spreadsheets now assists in reasoning over protein fold relationships. The generative AI stack is no longer a separate discipline from medicinal chemistry — it is medicinal chemistry's support layer, and it is already visibly reducing the time it takes to reach a candidate molecule entering the vivo testing phase.

7 Why All of This Matters Now

What is the larger narrative that threads these three tracks together? Not simply "tech is changing fast" — that is a self-referential statement. A more precise account goes like this: the world is confronting a period of "intelligence surplus," where the ability to make accurate predictions, decisions, and interventions is being inducted into systems across all three physical layers — digital, mechanical, and biological — simultaneously. The price per unit of inference is collapsing. The design cycle for a new system is shrinking. The trustworthiness of prediction at the frontier is high enough that betting significant corporate capital on it is a rational move.

This convergence places very specific demands on institutions that have historically been decision-brake technologies rather than accelerators. Regulatory regimes built for discrete, slow-advancing, individually inventoried technologies (a drug molecule, a chip design, an automotive safety standard) are now being asked to adapt their decision-making architecture to move at the same velocity as the technology they govern. Industry realises this. Enterprises that try to absorb the agentic AI wave, the autonomous-car stack, and biotech innovation simultaneously are in practice in the most demanding years of work this decade will ask of them.

For individuals — developers, founders, researchers, or simply curious observers — the message is one of meaningful leverage: the capability cost of building systems that exhibit intelligence, whether across software, cars, or gene-edits is dropping, and the window for building durable expertise while these fields solidify is narrowing. The frontier models released between late 2025 and early 2026 will likely look archaic again by late 2026. The autonomy stack that Rivian is rolling out across its R2 and R1 vehicles may look dated when Apple's Project Titan revisions arrive with their own chip strategies. The generative-biology tools assisting CRISPR design at biotech firms are themselves speeding up — every six months in 2025. That is the shape of normal now.

8 Looking Ahead

The three stories told here carry strong mutual reinforcement. The AI model competition is indirectly advancing biotech, because biotech researchers can now use frontier models to perform highly precise task reasoning without hiring entire ML teams for every new problem domain. The autonomous-driving sensor stack — particularly LiDAR, which requires significant signal-processing capacity — is an application that will drive edge-AI silicon forward in a direction that applies back to other edge contexts, including medical instruments. mRNA delivery platform development is a modular abstraction that future personalised medicines — including genetically targeted gene and cell therapies — will inherit directly. The threads are already knotted.

The most interesting unforced prediction is not who dominates which track, but who figures out how to own the integration layer. AI models that interface directly with biotech data structures, autonomous vehicle compute that manages its own training feedback loop, genetic editing pipelines that use AI to optimise their own design cycle — those integration stories will define the best returns of this decade. The companies that identify first where intelligence, autonomy, and biology overlap and build defensible infrastructure at those junctures will matter the most.

Sources

• OpenAI — Introducing GPT-5.4, March 5, 2026, openai.com
• OpenAI — Introducing GPT-5.2, December 11, 2025, openai.com
• OpenAI — Introducing GPT-5.5, April 23, 2026, openai.com
• OpenAI — Introducing GPT-5.1 Codex-Max, November 19, 2025, openai.com
• Google — Gemini 3: A New Era of Intelligence, November 18, 2025, blog.google
• Google DeepMind — Gemini 3.1 Pro Model Card, February 19, 2026, deepmind.google
• Anthropic — Introducing Claude 4, May 22, 2025, anthropic.com
• Anthropic — Introducing Claude Sonnet 4.5, September 29, 2025, anthropic.com
• Anthropic — Introducing Claude Sonnet 4.6, February 17, 2026, anthropic.com
• CNBC — Rivian Announces New AI Tech, Chip, and Robotaxi Ambitions, December 11, 2025, cnbc.com
• The Verge — Rivian Is Designing Its Own Powerful AI Chips for Autonomous Driving, December 11, 2025, theverge.com
• TechCrunch — Rivian Goes Big on Autonomy, with Custom Silicon, LiDAR, and a Hint at Robotaxis, December 11, 2025, techcrunch.com
• Science/AAAS — Gene-Editing Therapy Made in Just 6 Months Helps Baby with Life-Threatening Disease, December 2025, science.org
• Nature — Infant Receives the First Customized CRISPR Therapy, December 2025, nature.com
• Penn Medicine — World's First Patient Treated with Personalized CRISPR Gene Editing Therapy, pennmedicine.org
• Stanford Medicine — AI-Powered CRISPR Could Lead to Faster Gene Therapies, September 2025, med.stanford.edu
• BioPharm International — 5 Ways mRNA Is Revolutionizing Drug Development Beyond Vaccines, July 2, 2025, biopharminternational.com
• BioPharm International — AI, CRISPR, and mRNA Driving Biotech's Smartest Decade Yet, December 27, 2025, biopharminternational.com