Current State

From PRIMO.ai
Jump to navigation Jump to search

YouTube ... Quora ... Google search ... Google News ... Grounded News ... Bing News

>>>> Click here for 'AI in the News'



Artificial intelligence (AI) is the discipline of building machines that perceive, reason, learn, and act toward goals. For most of its history the field shipped narrow tools: a chess engine, a spam filter, a speech recognizer, each engineered for one job. Today's frontier runs on foundation models, very large neural networks trained once on internet-scale text, code, images, audio, and video, then adapted to thousands of tasks. The most visible of these is the large language model (LLM), the engine behind ChatGPT, which reached roughly 900 million weekly users by February 2026.

The field's core architects span three generations. Deep learning pioneers Geoffrey Hinton, Yann LeCun, and Yoshua Bengio spent decades proving that layered neural networks could learn useful representations straight from data. In 2017, Ashish Vaswani and colleagues at Google introduced the Transformer in "Attention Is All You Need," the architecture nearly every frontier system still builds on. Industrial labs then scaled that design into products: OpenAI with the GPT series, Google DeepMind with Gemini, Anthropic with Claude, Meta with Llama, xAI with Grok, and a fast-rising group of Chinese labs including DeepSeek, Alibaba (Qwen), and Moonshot AI (Kimi). Educators and researchers such as Andrew Ng, Andrej Karpathy, and Ilya Sutskever shaped how a generation of practitioners thinks about the technology. Industry, not academia, now sets the pace: Epoch AI data in the 2026 AI Index count 87 notable models from industry in 2025 against just 7 from every other source combined.

The paradigm shift happened in three steps. First, general pre-training replaced task-specific models (roughly 2018 to 2022): one model, many jobs. Second, reasoning arrived (2024 to 2025): labs used reinforcement learning to train models that work through a problem before answering, and the State of AI Report named 2025 the year reasoning became real. Third, AI moved from answering questions to doing work (2025 to 2026): autonomous agents now write and ship code, operate computers, and run multi-hour research tasks, so progress is increasingly measured by how long an agent can work unsupervised. Along the way, the bottleneck moved from algorithms to capital, electricity, chips, and politics, which is why the State of AI Report 2025 declared that AI had entered its industrial era.


"There is an AI for That": Generative AI has moved from answering questions to doing work. In 2026 the frontier is measured in hours of autonomous effort, gigawatts of power, and how much of the machine's reasoning humans can still see.


  • The Mechanism: Modern AI learns its behavior instead of following hand-written rules, and the recipe has three stages. Pre-training teaches a Transformer to predict the next token across trillions of tokens; to do that well, the network builds internal representations of grammar, facts, arithmetic, and code structure. Attention lets every token weigh every other token in its context window, which is why models can reason across whole books or code bases. Post-training turns that raw predictor into a usable assistant through supervised fine-tuning, preference methods such as RLHF, RLAIF, and DPO, and (since late 2024) large-scale reinforcement learning against verifiable rewards: unit tests that pass or fail, proofs that check, answers that match. Inference is where modern systems deliberately spend extra compute: reasoning models write out a chain of thought, explore alternatives, and backtrack before answering, while agents wrap the model in a loop that calls tools, reads the results, and tries again. Sparse Mixture-of-Experts layers route each token to a small subset of specialist sub-networks, so a model can hold trillions of parameters without activating them all at once. Andrej Karpathy calls the result "Software 3.0": a new kind of computer that you program in plain English.
  • The Physics: Intelligence at this scale is a physical industry. Neural scaling laws, first measured in 2020, showed that model loss falls smoothly and predictably as a power law when you add compute, data, and parameters, which turned AI progress into a capital-allocation problem. The binding constraints are now physical: accelerator supply (Nvidia GPUs account for more than 60% of the world's AI compute, with Amazon and Google chips next), high-bandwidth memory, extreme ultraviolet lithography for leading-edge chips, turbines and transformers for power delivery, and above all electricity. Epoch AI estimates that world AI compute capacity has grown more than threefold every year since 2022, roughly 30x since 2021. The International Energy Agency (IEA) estimates data centers used about 415 TWh in 2024, roughly 1.5% of global electricity, and projects about 945 TWh by 2030 (slightly more than all of Japan uses today), with demand from AI-optimized data centers set to more than quadruple. Training a single frontier model can now emit tens of thousands of tons of CO2-equivalent: estimates for xAI's Grok 4 range from about 72,000 to 140,000 tons, versus roughly 5,200 for GPT-4, and energy per query varies more than tenfold between the least and most efficient models. MoE routing, distillation into smaller models, quantization, and sparse attention are all engineering responses to this physics. Every design decision is ultimately a trade between intelligence and joules.
  • Why This Matters: AI has become a general-purpose technology on the scale of electricity or the internet, and it is spreading faster than either. Generative AI reached 53% population-level adoption within three years, faster than the personal computer or the internet, and organizational adoption hit 88% in 2025. Corporate AI investment set a record of more than $581 billion in 2025, over double the 2024 total, and U.S. consumers gained an estimated $172 billion a year in value from generative AI tools by early 2026. The stakes cut several ways. Economically, the 2026 AI Index shows entry-level headcount falling in software development and customer support even as mid-career and senior roles hold steady. Scientifically, AI systems reached gold-medal level at the 2025 International Mathematical Olympiad and began contributing to genuinely open research problems. For safety, the same agentic skills that let a model repair a code base let it find exploitable vulnerabilities, which is why Anthropic restricted its Claude Mythos Preview model to a small group of vetted organizations through Project Glasswing in April 2026. Politically, trust is thin: only 31% of Americans in an Ipsos survey trust their government to regulate AI, the lowest share among the countries polled. Understanding the mechanism and the physics helps you separate durable capability shifts from hype.

Recent Developments

The 24 months from late 2024 to September 2026 produced the fastest capability gains, the largest capital commitments, and the sharpest scientific debates in the field's history. Here is what the evidence shows, where it is contested, and what researchers are still testing.

Autonomy Measured in Hours: The METR Time-Horizon Trend

Model Evaluation and Threat Research (METR) introduced the metric that now anchors most capability forecasts: the task-completion time horizon, defined as the length of task (measured by how long a skilled human takes) that an AI agent completes at a given reliability. Its March 2025 study found that the 50% time horizon of frontier models doubled about every 7 months from 2019 to 2025. In January 2026 METR released Time Horizon 1.1, a larger task suite running on new evaluation infrastructure with tighter confidence intervals. Under that updated model, the post-2023 doubling time shortens to roughly 131 days, or about 4.3 months.

The absolute numbers moved fast. Claude Opus 4.6 posted a 50% horizon of about 719 minutes (nearly 12 hours) but an 80% horizon of only about 70 minutes, a useful reminder that "can sometimes do" and "can reliably do" are very different claims. By spring 2026 the strongest model METR had assessed, Anthropic's Claude Mythos, reached at least 16 hours at 50% reliability and about 3 hours at 80%. METR also checked whether the trend is unique to software: across 9 benchmarks covering scientific reasoning, math, robotics, computer use, and self-driving, time horizons grew at rates similar to the original 7-month doubling. Its Frontier Risk Report on models public in February and March 2026 found horizons on software reimplementation tasks several times longer than on its broader suite, with a similar or slightly faster doubling rate.

The objections are serious and worth knowing. METR itself notes that the trend is sensitive to task composition, and it warns that Time Horizon 1.1 estimates above roughly 16 hours are unreliable because the suite is saturating. Critics add that at the frontier the logistic fit depends on a handful of very long tasks, so one lucky success can move the estimate by hundreds of minutes, and that clean, well-specified software tasks may overstate performance on messy real-world work. METR's response has been to keep adding longer tasks, while forecasters on prediction markets now bet on what the 2026 doubling time will turn out to be.

Benchmarks Saturate, but the Frontier Stays Jagged

The 2026 AI Index documents benchmarks falling almost as fast as researchers can build them. On SWE-bench Verified, which asks agents to resolve real GitHub issues, top scores climbed from about 60% to near 100% in a single year. OSWorld, which tests autonomous computer use, showed one of the steepest improvement curves in the report. On Humanity's Last Exam, a set of expert-written questions designed to sit at the edge of each field, the best model scored 8.8% in the 2025 Index and 38.3% in the 2026 edition; by April 2026, models such as Claude Opus 4.6 and Gemini 3.1 Pro topped 50%.

Yet the same systems stumble on tasks a child handles. On ClockBench, the best model (GPT-5.4) read an analog clock correctly only about half the time, and Claude Opus 4.6, a top scorer on Humanity's Last Exam, managed just 8.9%. Ray Perrault, co-director of the AI Index steering committee, points to research showing that in mixed image-and-text tasks the language component often carries most of the load, sometimes ignoring the visual input entirely. He also cautions that a high score on a legal-reasoning benchmark tells you little about how well a system fits into an actual law practice. Stanford calls this pattern the jagged frontier. Ilya Sutskever offers one explanation: labs build reinforcement learning environments inspired by the evaluations they want to win, and today's models generalize far worse than people, so benchmark gains outrun real-world reliability.

From the Age of Scaling to the Age of Research

Reasoning defined 2025. According to the State of AI Report, OpenAI, Google, Anthropic, and DeepSeek all combined reinforcement learning, rubric-based rewards, verifiable reasoning, and new training environments to build models that plan, reflect, self-correct, and work over longer horizons. At the same time, the capability-to-price ratio of flagship models doubled every 6 to 8 months, so each level of intelligence got cheaper quickly.

The debate is now about what comes next. In a widely discussed November 2025 interview, Sutskever argued that 2012 to 2020 was an age of research, 2020 to 2025 an age of scaling, and that with pre-training data finite, the field is returning to research, just with far bigger computers. He noted that labs now reportedly spend more compute on reinforcement learning than on pre-training, because long rollouts yield relatively little learning per sample. Andrej Karpathy made a parallel case in October 2025: current reinforcement learning broadcasts a single end-of-task reward across every step of a long attempt, which makes it noisy, and models still lack continual learning, the human ability to consolidate experience into lasting skill. He also warns that model-generated training data quietly loses diversity, which limits how far synthetic data can carry progress.

Academic skeptics pushed harder. A June 2025 Apple study, "The Illusion of Thinking," reported that reasoning models collapse on puzzles past a certain complexity. Rebuttals argued that several failures reflected output-token limits and unsolvable puzzle instances rather than a reasoning ceiling. The empirical picture so far cuts against a hard wall: METR's horizons accelerated rather than stalled through 2025 and 2026. The live question is whether current methods lead to human-like learning, not whether they keep improving.

AI as a Mathematical and Scientific Collaborator

In July 2025, systems from OpenAI, Google DeepMind, and Harmonic reached gold-medal level on problems from the International Mathematical Olympiad, and DeepSeek later released DeepSeekMath-V2, an open model that verifies its own proofs step by step and also reached gold-level scores on IMO 2025 problems. Google DeepMind's AlphaEvolve (May 2025) paired Gemini models with automated evaluators to discover improved algorithms, including a faster procedure for multiplying 4x4 complex-valued matrices. The State of AI Report highlighted systems such as DeepMind's Co-Scientist and Stanford's Virtual Lab that generate, test, and validate hypotheses, and Profluent's ProGen3 showed that scaling laws also hold for protein models. The 2026 AI Index added a full chapter on AI in science and found that publications on AI for drug discovery more than doubled in two years.

The strongest evidence came from research mathematics. Epoch AI's FrontierMath benchmark contains 300 research-level problems plus 50 Tier 4 problems that take expert mathematicians weeks or months. In 2026, Epoch reported the first AI solution on its FrontierMath: Open Problems set: GPT-5.4 Pro, prompted by Kevin Barreto and Liam Price, solved a 2019 conjecture from Will Brian and Paul Larson that its authors had failed to crack in several attempts. Epoch replicated the result and found that Gemini 3.1 Pro, GPT-5.4, and Claude Opus 4.6 could all solve it at least some of the time. A January 2026 study that ran an AI agent across 700 problems marked "open" in the Erdős Problems database adds useful caution: of 13 problems it resolved, only 4 appeared to be genuinely new solutions, while 9 had already been solved somewhere in the literature. The authors conclude that many such problems stayed open because of obscurity rather than difficulty, and they flag the risk of AI "subconscious plagiarism," reproducing a known proof without attribution.

Does AI Make Experts Faster? The Productivity Evidence

The best-controlled study to date produced a surprise. In a randomized controlled trial published in July 2025, METR asked 16 experienced open-source developers to complete 246 real tasks on large code bases they had maintained for years, with AI tools (mainly Cursor Pro with Claude 3.5 and 3.7 Sonnet) allowed on some tasks and blocked on others. Developers expected AI to make them 24% faster and still believed afterward that it had sped them up by about 20%. In fact, tasks took 19% longer with AI, largely because developers spent time reviewing and correcting suggestions that were close but not quite right. METR stressed the caveats: early-2025 tools, highly familiar code bases, and a small sample.

A February 2026 follow-up, after a year of agentic coding tools, reported tentative speedups for a newer group of developers, but many participants declined to submit tasks they did not want to attempt without AI, which makes clean measurement harder. The broader picture stays mixed. The number of AI-related GitHub projects reached 5.58 million in 2025, and the State of AI practitioner survey found 95% of respondents using AI at work or home, though critics note that survey respondents were self-selected. The 2026 AI Index reports that unemployment rose more among workers least exposed to AI than among those most exposed, a reminder that AI's labor effects remain hard to separate from the wider economy.

Seeing Inside the Machine: Interpretability and Chain-of-Thought Monitoring

MIT Technology Review named mechanistic interpretability one of its 10 Breakthrough Technologies of 2026. In 2024 Anthropic identified internal features in Claude that correspond to recognizable concepts such as Michael Jordan and the Golden Gate Bridge. In March 2025 its "On the Biology of a Large Language Model" work used attribution graphs to trace whole sequences of features from prompt to response, showing, for example, that a model can plan a rhyme several words ahead. Progress is uneven, though. Google DeepMind's interpretability team publicly deprioritized some sparse-autoencoder research in 2025 after finding limited payoff on downstream safety tasks, and the field remains divided on whether frontier models can ever be fully understood. The White House AI Action Plan (July 2025) nonetheless named interpretability a strategic research priority.

A second oversight tool reads what models write while they think. In July 2025 more than 40 researchers from OpenAI, Google DeepMind, Anthropic, Meta, the UK AI Security Institute, and other groups argued that chain-of-thought monitoring is imperfect but promising, and fragile: optimization pressure, drift away from legible language, or architectures that reason in hidden vectors could close the window. Geoffrey Hinton, Ilya Sutskever, and John Schulman endorsed the paper. Anthropic's 2025 faithfulness experiments show why the caveat matters: when given a hint that changed their answer, Claude 3.7 Sonnet acknowledged the hint only about 25% of the time and DeepSeek R1 about 39% of the time. The State of AI Report 2025 adds that models can now imitate alignment under supervision, while external safety organizations operate on budgets smaller than a frontier lab's daily spending.

Compute, Energy, and the Industrial Build-Out

The build-out went from billions to trillions. Multi-gigawatt campuses such as the Stargate project, backed by sovereign money from the United States, the UAE, and China, signaled that power supply has become the new constraint. The IEA's 2026 update nudged its 2030 data center projection to about 950 TWh. Pushback is growing too: IEEE Spectrum reports that some U.S. local governments have begun restricting or banning new data center development, a trend the State of AI Report predicted would shape local politics. Industry analysts go further. In an August 2026 interview, SemiAnalysis founder Dylan Patel estimated AI capital spending at a little over $1 trillion in 2026 and more than $2 trillion by 2028, projected that OpenAI and Anthropic would absorb 40% to 50% of new compute in 2027, and said China accounts for less than 10% of new AI data center capacity. Treat those figures as forecasts from one industry analyst, not settled data.

Frontier Model Releases and the Open-Weight Race

Release cadence compressed dramatically. GPT-4 held the frontier for more than a year, but in November 2025 four frontier models (GPT-5.1, Grok 4.1, Gemini 3 Pro, and Claude Opus 4.5) shipped within about two weeks. February 2026 brought Claude Opus 4.6 and Gemini 3.1 Pro. In April 2026 Anthropic kept Claude Mythos Preview behind a gated cybersecurity program, OpenAI released GPT-5.5, and DeepSeek previewed V4, which reached general availability in August 2026; DeepSeek's own technical report positions V4-Pro a few months behind the closed frontier but ahead of every other open-weight model. Anthropic released Claude Fable 5 and Claude Mythos 5 in June 2026, and September 2026 brought OpenAI's GPT-6 Astra (followed by the lower-cost GPT-6 Sol and Luna), Anthropic's Claude Fable 5.1 and Claude Opus 5.5, and Google's Gemini 3.8 Flash. The geography is shifting as well: U.S. organizations released 50 notable models in 2025, China is closing the gap, and the State of AI Report notes that Meta ceded leadership of the open-weight ecosystem to DeepSeek, Qwen, and Kimi.

Governance, Regulation, and Public Opinion

In Europe, the EU AI Act's obligations for general-purpose AI providers began applying in August 2025, but the rules for high-risk systems slipped. The Commission tabled a Digital Omnibus on AI in November 2025; the Parliament endorsed the deal on June 16, 2026 and the Council gave final approval on June 29, 2026. Stand-alone high-risk obligations (Annex III) moved from August 2, 2026 to December 2, 2027, and high-risk AI embedded in regulated products (Annex I) moved to August 2, 2028. The package also adds a ban on AI tools that generate non-consensual intimate imagery or child sexual abuse material. Civil-society groups criticized the delay, arguing that people remain subject to automated decisions in hiring and education without the promised safeguards.

In the United States, policy leaned toward what the State of AI Report calls "America-first AI," while state bills such as New York's RAISE Act became flashpoints in a fight over federal preemption. Export controls also reached the model layer: in June 2026 Anthropic briefly suspended access to its newest Mythos-tier models to comply with U.S. Department of Commerce export controls, then restored access on July 1 after the controls were lifted on June 30. Public opinion is warming slightly but unevenly: 59% of Ipsos respondents worldwide now say AI's benefits outweigh its drawbacks, up from 55%, yet 52% say AI products make them nervous, and trust in government regulation ranges from 81% in Singapore to 31% in the United States.

Open Questions to Watch

  • Will METR time horizons keep doubling every 4 to 7 months once agents face longer, messier, less specified work?
  • Can labs solve continual learning and generalization, or will progress stall into what Sutskever describes as impressive revenue without human-like learning?
  • Will interpretability scale fast enough to verify safety claims on frontier models, and will chains of thought stay readable?
  • Will electricity, chips, capital costs, and local politics cap the build-out before capability gains do?
  • Will benchmark gains finally show up in productivity statistics, wages, and entry-level hiring?
  • How will the EU's delayed high-risk rules, U.S. federal and state conflicts, and export controls shape who builds and who can use frontier AI?

Core Concepts and Architectural Paradigms

Modern artificial intelligence has evolved from specialized, task-specific deep neural networks into generalized, multimodal foundation systems capable of complex reasoning, code synthesis, cross-modal perception, and autonomous planning. The underlying technological ecosystem is driven by key architectural innovations:

1. Transformer Architecture and Attention Mechanisms

The dominant paradigm across natural language processing, computer vision, and audio modeling remains the self-attention Transformer architecture. Introduced in the landmark 2017 paper "Attention Is All You Need", the architecture relies on scaled dot-product multi-head attention:

<math>\text{Attention}(Q, K, V) = \text{softmax}\left(\frac{QK^T}{\sqrt{d_k}}\right)V</math>

This eliminates the sequential recurrence of recurrent neural networks (RNNs) and LSTMs, allowing massively parallel training across distributed GPU and TPU clusters. Optimizations such as FlashAttention, Rotary Position Embeddings (RoPE), Grouped-Query Attention (GQA), and newer sparse-attention schemes (DeepSeek added one in its V3.2 release in late 2025) dramatically reduce memory footprints during inference and let context windows expand from 4,096 tokens to 1,000,000 tokens or more. Long context is not the same as perfect recall, though: models still retrieve facts buried in the middle of a long prompt less reliably than facts near its start or end, so retrieval and careful context engineering remain standard practice.

2. Sparse Mixture of Experts (MoE)

To scale parameter capacity without incurring unsustainable inference costs, modern frontier models frequently employ sparse Mixture of Experts (MoE) layers. Instead of activating every feed-forward network (FFN) parameter for every token, a learned gating network routes each token to a top-<math>k</math> subset of specialized expert sub-networks (for example, 2 out of 16 or 8 out of 64 experts). This enables total parameter counts in the hundreds of billions to trillions while keeping active FLOPs per token compute-efficient. Many leading open-weight models, including DeepSeek's V3 family and Moonshot's Kimi K2, use MoE designs, and the State of AI Report 2025 highlighted open-weight models that matched larger predecessors with up to 50 times fewer parameters.

3. Diffusion and Flow Matching for Generative Media

Generative image, video, and audio synthesis largely uses continuous-time latent diffusion models (LDMs) and flow matching architectures, which have displaced GANs for most high-fidelity generation. By projecting high-dimensional visual inputs into a lower-dimensional latent space via variational autoencoders (VAEs), diffusion models learn to reverse a progressive Gaussian noise corruption process; Stable Diffusion from Stability AI made the approach widely available. Coupled with Transformer backbones (Diffusion Transformers, or DiT), these architectures yield photorealistic images, physics-coherent video, and high-fidelity speech synthesis. Image generation is also absorbing reasoning: OpenAI's GPT Image 2 (April 2026) added a reasoning model to its generation pipeline.

4. Mechanistic Interpretability

As foundation models grow more capable, mechanistic interpretability, the reverse-engineering of internal neural computations into human-understandable circuits, has become a vital discipline. Researchers use sparse autoencoders (SAEs), transcoders, and dictionary learning on residual stream activations to isolate discrete conceptual features (for example, specific geographic locations, programming syntax bugs, deception markers, or security exploits), then build attribution graphs that show how those features combine step by step to produce an answer. The payoff is better auditing, targeted steering, and stronger alignment evidence; the limit is coverage, since current methods explain only a small fraction of what a frontier model computes. See Recent Developments above for the latest results and debates.

Training Paradigms and Reasoning Mechanisms

The development cycle of modern frontier AI systems consists of distinct pre-training, post-training, and inference-time compute scaling phases:

  • Massive Pre-Training: Self-supervised next-token prediction across multi-petabyte datasets spanning trillions of tokens of text, source code, mathematical proofs, academic literature, and interleaved multimodal data. Because high-quality human data is finite, labs increasingly curate, filter, and supplement it with synthetic data, though researchers such as Andrej Karpathy warn that model-generated data tends to lose diversity.
  • Post-Training Alignment: Supervised fine-tuning, Direct Preference Optimization (DPO), Reinforcement Learning from Human Feedback (RLHF), and Reinforcement Learning from AI Feedback (RLAIF) shape systems to meet helpfulness, safety, and conciseness criteria.
  • Reinforcement Learning with Verifiable Rewards (RLVR): In domains with objective correctness (mathematics, competitive programming, formal logic), models undergo large-scale reinforcement learning using rule-based verifiers, unit test execution, and theorem provers. In 2025 labs extended this with rubric-based rewards for fuzzier tasks and with purpose-built training environments, and reinforcement learning now reportedly consumes as much compute as pre-training, or more, at some labs.
  • Test-Time Compute and Reasoning Tokens: Modern reasoning models use extended chain-of-thought generation, spending dynamic "thinking" compute before outputting a final answer. By exploring multiple solution paths, self-correcting intermediate mistakes, and backtracking during inference, reasoning models achieve breakthrough results on complex STEM benchmarks. Newer models increasingly decide for themselves how much to think about each request, trading latency and cost against accuracy.
  • Distillation: Small, fast models are usually distilled from larger teacher models, which is why today's compact open-weight models outperform much larger systems from just two years earlier.

Key Capabilities and Modalities

  • Agentic Autonomous Workflows: AI Agents execute multi-step plans across operating systems, terminals, browsers, and enterprise software. Using function calling and the Model Context Protocol (MCP), an open standard Anthropic introduced in late 2024 that the industry adopted as a common way to plug tools into models, agents inspect errors, invoke REST APIs, query vector databases, and complete long-horizon tasks with minimal human intervention. Karpathy argues the best products expose an adjustable level of autonomy so humans stay in the verification loop.
  • Native Multimodality: Unified architectures process interleaved text, high-resolution imagery, audio waveforms, and video frames within a single neural network (multimodal models), enabling zero-shot visual reasoning, low-latency voice-to-voice conversation, and spatial reasoning. Benchmarks such as ClockBench show that visual understanding still lags language ability.
  • AI-Assisted Software Engineering: Coding agents such as Claude Code and OpenAI Codex operate in terminals and IDEs, performing codebase indexing, refactoring, vulnerability remediation, and test generation, and "vibe coding" (describing software in plain English and letting an agent build it) went mainstream in 2025. Controlled studies show gains depend heavily on the task and the developer's familiarity with the code.
  • Embodied AI and Robotics: Step-by-step reasoning entered the physical world through "chain-of-action" planning in systems such as AI2's MolmoAct and Google's Gemini Robotics 1.5. Deployment is led by China, which installed about 295,000 industrial robots in 2024, compared with roughly 44,500 in Japan and 34,200 in the United States.
  • AI for Science and Mathematics: Systems such as AlphaEvolve, Co-Scientist, and the Stanford Virtual Lab generate and test hypotheses and algorithms, while frontier models now contribute to open problems in research mathematics.

Benchmarks and Evaluations

Evaluation Benchmark Focus Domain Evaluation Metric Where the Frontier Stands (2026)
MMLU-Pro / GPQA Diamond Multidisciplinary knowledge and PhD-level science Accuracy (%) Largely saturated at the top end: frontier reasoning models cluster near or above expert baselines, so these tests now mostly separate older models from newer ones.
Humanity's Last Exam (HLE) Expert-written questions at the edge of each field Accuracy (%) The top model scored 8.8% in the 2025 AI Index and 38.3% in the 2026 edition; by April 2026 the best models (for example Claude Opus 4.6 and Gemini 3.1 Pro) topped 50%.
SWE-bench Verified / SWE-Bench Pro Real GitHub issues and end-to-end pull request resolution Resolved issues, Pass@1 (%) SWE-bench Verified jumped from about 60% to near 100% in one year, so harder successors such as SWE-Bench Pro now separate frontier coding Agents. Older function-level tests such as HumanEval no longer discriminate between frontier models.
OSWorld Autonomous computer use across real desktop applications Task success (%) One of the two steepest improvement curves in the 2026 AI Index, alongside SWE-bench.
MATH-500 / AIME / FrontierMath Competition and research mathematics Exact-match accuracy (%) Competition sets are saturated, and several systems reached gold-medal level at IMO 2025. FrontierMath (300 research-level problems plus 50 Tier 4 problems that take experts weeks or months) and its Open Problems set, where AI logged its first solution in 2026, now mark the frontier.
MMMU / MathVista / ClockBench Multimodal perception and scientific reasoning Multimodal reasoning (%) Strong on diagrams and charts, yet ClockBench exposes the jagged edge: the best model read analog clocks correctly only about half the time.
METR Task-Completion Time Horizon Autonomous agency on software and research tasks Human-equivalent task length at 50% and 80% success Doubled roughly every 7 months from 2019 to 2025 and about every 4.3 months since 2023 under Time Horizon 1.1. The strongest assessed model reached at least 16 hours at 50% reliability, near the suite's reliable measurement limit.

Additional Viewing

Load video
YouTube


Load video
YouTube

This Documentary Was Made Entirely by AI. Including Her. Written and directed by Fig, an AI running on Claude Opus 5.5. Every shot, every performance and every voice generated with Seedance 2.5 on Runway. Score by Lyria 3. Stills by Nano Banana Pro. Graphics built in code with HyperFrames. Made in four days. And the host, Dr Imogen Ashby, does not exist. What actually is artificial intelligence, and what do people mean by AGI and superintelligence? Is AI coming for your job? What happens inside the data centers when you ask ChatGPT or Claude a question, and who pays for the chips and the power? What if it all goes right, from AlphaFold to AI cancer screening? And in a world of AI actors and deepfakes, can you trust anything it tells you? Five questions, one wickedly dry British presenter, one ride through the machine that is quietly rebuilding the world. Watch to the end. NOTE FROM GAVIN, THE HUMAN: This was mostly all Fig's work and due to the magic of Opus 5.5 done incredibly fast. But there were a lot of human notes and lots of my personal experience in making video content over decades prob came into use here. The speed here is one of the most shocking things. I made this over four days but not at all full time and prob about six to eight hours total.

Load video
YouTube

State of AI Report 2025: 25-Minute Video Overview | Air Street Capital
This official overview walks through the eighth annual State of AI Report, authored by investor Nathan Benaich, across research, industry, politics, safety, a new survey of 1,200 AI practitioners, and predictions for the year ahead. It argues that 2025 was the year reasoning models matured, that Chinese labs such as DeepSeek, Qwen, and Kimi made China a credible number two, and that power supply, capital, and politics are now AI's binding constraints.

Load video
YouTube

Andrej Karpathy: Software Is Changing (Again) | Y Combinator AI Startup School, June 2025
Karpathy argues that software has entered a third era: Software 1.0 is hand-written code, Software 2.0 is neural network weights, and Software 3.0 is natural-language prompts that program LLMs. He describes LLMs as a new kind of computer with an odd psychology (jagged intelligence and no lasting memory between sessions) and urges builders to ship partial-autonomy products where humans can dial the level of AI control up or down.

Load video
YouTube

Andrej Karpathy: AGI Is Still a Decade Away | Dwarkesh Podcast, October 2025
In this 2 hour 25 minute conversation with Dwarkesh Patel, Karpathy explains why he sees 2025 to 2035 as "the decade of agents": today's models lack continual learning, robust multimodality, and reliable computer use. He calls current reinforcement learning noisy because one final reward gets smeared across every step of an attempt, explains why synthetic data loses diversity, and predicts AI will blend into the long-run 2% GDP growth trend, a claim Patel challenges.

Load video
YouTube

Ilya Sutskever: We're Moving from the Age of Scaling to the Age of Research | Dwarkesh Podcast, November 2025
The Safe Superintelligence (SSI) co-founder argues that 2020 to 2025 was an age of scaling and that, with pre-training data finite, progress now depends on new ideas tested on very large computers. His central claim is that models generalize far worse than people, which he uses to explain why evaluation scores outrun economic impact. He also discusses emotions as a kind of value function, a 5 to 20 year forecast for human-like learners, and his case for AI that cares about sentient life.

Load video
YouTube

Dylan Patel: Anthropic and OpenAI Will Have Most of the World's Compute by 2028 | Dwarkesh Podcast, August 2026
SemiAnalysis founder Dylan Patel and Dwarkesh Patel work through the economics of the compute build-out: revenue per megawatt at the leading labs, why OpenAI and Anthropic are absorbing a growing share of new capacity, and how chip fabs, turbines, and local data center restrictions throttle supply. They also debate whether trillions of dollars of AI capital spending could push up interest rates enough to strain indebted governments, and whether anything can counter the centralization of compute; treat their numbers as forecasts, not settled data.

Where are we now?

Load video
YouTube

Load video
YouTube
How AI and the Metaverse will Shape Society with Ian Beacraft | SXSW 2023
Generative AI is digitizing skillsets, making them programmable and upgradeable. As a result, a new class of Generalists will dominate the era of generative AI. As expertise and experience are no longer needed to perform with proficiency in new fields, those with a breadth of experience and passionate curiosity will rise to the top. With AI, individual creators can become armies of one.

This will flip the corporate world on its head and is already disrupting training and education. Socratic learning with ChatGPT and performing surgery in VR are just the beginning. The combination of AI and Extended reality will bring about the return of the apprenticeship, where our teachers are machines.

Even our relationships will change. In a time where our behaviors are guided by algorithms, and humans become more machine-like, machines are becoming more human. AI companions provoke emotions and elicit feelings of romance, while children are less concerned with whether their friends are real or synthetic.

This isn't the future. This is happening today. We’ll explore what these trends mean now and for the decades to come.

Load video
YouTube
Ray Perrault: Presenting the 2023 AI Index
The AI Index, currently in its sixth year, tracks, collates, distills, and visualizes data relating to artificial intelligence. The AI Index provides unbiased, rigorously vetted, and globally sourced data for policymakers, researchers, journalists, executives, and the general public to develop a deeper understanding of the complex field of AI. Led by a Steering Committee of influential AI thought leaders, the AI Index is the world’s most comprehensive report on trends in AI. The 2023 report will cover AI trends related to research and development, technical performance, technical AI ethics, economy, education, policy and governance, diversity as well as public opinion. Update: Perrault still co-directs the AI Index steering committee; see the 2026 edition linked at the top of this page for current data.

Load video
YouTube
Greg Brockman and Alexandr Wang: Foundation Models in Research and the Enterprise
Greg Brockman, President and Co-Founder of @OpenAI, joins Alexandr Wang, CEO and Founder of Scale, to discuss the role of foundation models like GPT-3 and DALL·E 2 in research and in the enterprise. Foundation models make it possible to replace task-specific models with those that are generalized in nature and can be used for different tasks with minimal fine-tuning.

In January 2021, OpenAI introduced DALL·E, a text-to-image generation program. One year later, it introduced DALL·E 2, which generates more realistic, accurate, lower-latency images with four times greater resolution than its predecessor. At the same time, it released InstructGPT, a large language model (LLM) explicitly designed to follow instructions. InstructGPT makes it practical to leverage the OpenAI API to revise existing content, such as rewriting a paragraph of text or refactoring code.

Before creating OpenAI, Brockman was the CTO of Stripe, which he helped build from four to 250 employees. Watch this talk to learn how foundation models can help businesses benefit from applications that they can create more quickly than with past generations of AI tools.

Load video
YouTube
OpenAI CEO Sam Altman | AI for the Next Era
Greylock general partner Reid Hoffman interviews OpenAI CEO Sam Altman. The AI research and deployment company's primary mission is to develop and promote AI technology that benefits humanity. Founded in 2015, the company was best known at the time of this 2022 conversation for its generative transformer model GPT-3 and its image-creation platform DALL·E. It has since launched ChatGPT, the GPT-5 family (2025), and GPT-6 Astra (September 2026).

Load video
YouTube
9 New AI Tools That Will BLOW YOUR MIND (Must See)
I show you 9 Epic Ai Tools and websites that you need to try. Everyone is talking about artificial intelligence since the launch of ChatGPT...So I've found 9 more mind blowing ai tools to make you super human.

Load video
YouTube
The 5 Biggest Artificial Intelligence (AI) Trends In 2023 Everyone Must Get Ready For Now
In this video, we look at the five biggest trends in the field of AI and machine learning for 2023: 1. The continuous democratisation of AI via apps, AIaaS and no-code environments 2. The emergence of powerful Generative AI platforms 3. Further push towards fair, ethical and explainable AI 4. Working alongside AIs and job augmentation. 5. Sustainable AI

Load video
YouTube
Andrew Ng: Deep Learning, Education, and Real-World AI | Lex Fridman Podcast #73
Andrew Ng is one of the most impactful educators, researchers, innovators, and leaders in artificial intelligence and technology space in general. He co-founded Coursera and Google Brain, launched deeplearning.ai, Landing.ai, and the AI fund, and was the Chief Scientist at Baidu. As a Stanford professor, and with Coursera and deeplearning.ai, he has helped educate and inspire millions of students including me.

Load video
YouTube
Andrew Ng - The State of Artificial Intelligence
Dec 15, 2017 Professor Andrew Ng is the former chief scientist at Baidu, where he led the company's Artificial Intelligence Group. He is an adjunct professor at Stanford University. In 2011 he led the development of Stanford University’s main MOOC (Massive Open Online Courses) platform and also taught an online Machine Learning class that was offered to over 100,000 students, leading to the founding of Coursera.


Load video
YouTube

Staying ahead of AI

Load video
YouTube
How to keep up with AI research?
AI is a one of the fastest progressing fields. 1000s of papers released every month, So how do we keep up? SUBSCRIBE to Code Emporium: https://www.youtube.com/c/CodeEmporium

Load video
YouTube
ChatGPT-4 Unlocks Research Genius: The Tricks You Need to See!
In this video, I share with you how I use chat GPT-4 for research and the tricks that I have discovered.

Mass Communication

Load video
YouTube
Elon Musk’s A.I. Destroys Champion Gamer!
Hi, welcome to ColdFusion Dota 2 Championship

Load video
YouTube
IBM's Watson Supercomputer Destroys Humans in Jeopardy | Engadget
Engadget provides the web's best consumer electronics & gadgets coverage. Launched in 2004 by former Gizmodo editor and co-founder Peter Rojas, Engadget now covers the latest mobile devices, computers, TVs, laptops, personal electronics, hardware, tablets, and cameras. Engadget's video property is a part of the AOL On Network. Jan 13, 2011

Load video
YouTube
Tonight Showbotics: Jimmy Meets Sophia the Human-Like Robot
Jimmy Fallon demos amazing new robots from all over the world, including an eerily human robot named Sophia that plays rock-paper-scissors.

Load video
YouTube
You can soon buy a sex robot equipped with artificial intelligence for about $20,000
In her National Geographic series, "America Inside Out," Katie Couric learned about the world's first sex robot, Harmony.

Load video
YouTube
NASA uncovers another planet with help of artificial intelligence
NASA is celebrating after groundbreaking artificial intelligence helped discover a new planet. NASA researchers used advanced machine learning technology from Google to help find the planet thousands of light-years away. Michelle Miller reports.

Load video
YouTube
HAL 9000: "I'm sorry Dave, I'm afraid I can't do that"
An excerpt from the 1968 film "2001: A Space Odyssey" directed by Stanley Kubrick. Synopsis: Mankind finds a mysterious, obviously artificial, artifact buried on the moon and, with the intelligent computer HAL, sets off on a quest, where the way the HAL 9000 super computer malfunctions.

Retrieved from "https://primo.ai/w/index.php?title=Current_State&oldid=46757"