8 sectors · 35 sub-themes Explore how AI is moving — sectors branch into sub-themes, and shared entities link them across the map. Click a sector to read its sub-themes, each claim traceable to a numbered source.
◍ Each cluster is a sector of the AI landscape — bigger stars are better-sourced sub-themes, and grey threads are entities that recur across sectors. Drag to pan, scroll to zoom.
Fresh evidence keeps chipping away at the premise that chain-of-thought reasoning is real reasoning. Researchers at IIIT Allahabad and NIELIT Delhi tested a wide range of proprietary and open models, including ChatGPT and Claude, and found the step-by-step traces often don't reflect what actually drives the answer. That echoes earlier Anthropic and Apple work on unfaithful reasoning, a debate that has since dissolved into dueling rebuttals. The response is to move reasoning off the page. A new survey argues that externalizing every step as text is wasteful and pushes latent, non-text reasoning. Chelsea Finn and collaborators go further, using RL to optimize the reasoning strategy itself, running parallel trains of thought and aggregating them.
Agentic coding has crossed into the mainstream: Anthropic reports the share of GitHub projects with coding-agent activity has more than doubled since late 2025, with Claude Code driving most of that growth. Claude Code creator Boris Cherny says he no longer writes his own code. The frontier problem is now duration — an Anthropic engineer ran a 70-minute workshop on keeping agents on-task for hours or days without looping, drifting, or losing context. Underpinning this is the fast standardization around Model Context Protocol. Microsoft (Dynamics 365 Commerce), X, Navan, and Propel all launched MCP servers, exposing enterprise data to agents through one interface. That connective tissue is what makes long-running agents useful across systems. The craft is shifting accordingly. Practitioners are floating "loop engineering" as prompt engineering's successor: stop hand-crafting one-off prompts, start designing persistent, self-supervising agent workflows. SaaStr notes even mediocre prompts now yield decent software, lowering the value of precise manual prompting. The through-line: attention is moving from single turns to durable, tool-connected loops.
ByteDance's Seedance 2.5 is the headline: native 30-second single-clip 4K video with no stitching, and reports of up to three-minute generations. General release is targeted for early July, and the model is already flowing into CapCut and Dreamina via the faster, cheaper Seedance 2.0 Mini. The through-line this cycle is a race on length, resolution, and cost. xAI's Grok Imagine Video 1.5 attacks the last axis hardest. It topped an AI video leaderboard while priced 86% below Sora, generating clips in roughly 25 seconds with sharper physics and better audio. Its image counterpart, Grok Imagine 1.5, hit wide release. Document AI saw its own leap. Mistral OCR 4 outputs structured data with bounding boxes, block classifications, and word-level confidence scores. It topped OlmOCRBench at 85.20 and beat AWS Textract, Azure Document Intelligence, and Gemini in blind human evaluation across 600+ documents. It's aimed squarely at RAG pipelines and agent-driven invoice and form processing. Midjourney went the strangest direction, standing up a "Midjourney Medical" division and a "Scanner" product, an unexpected move into imaging hardware alongside v8.1's better moodboards. Notably, creators are now chaining these tools together. Emerging workflows combine Midjourney, GPT Image 2, Seedance 2.0, Suno, and Nano Banana Pro in single pipelines. The models are converging into interoperable production stacks, not standalone toys.
OpenAI moved to build its own silicon. With Broadcom and Celestica, it unveiled "Jalapeño," an inference chip designed around OpenAI's own LLM workloads, not a general accelerator. Hock Tan confirmed first silicon samples are in hand and under test, with deployment expected at Microsoft. The intent is explicit: reduce dependence on Nvidia and own the full stack. Compute scarcity is now the industry's binding constraint. Google capped Meta's Gemini access — a raw capacity shortfall, not a pricing fight. Meta's answer was to launch "Meta Compute" to rent out its own excess capacity, sending the stock up roughly 9% while analysts warned of margin dilution. SK Telecom went bigger still, committing ~$653 billion to 15 GW of Korean data-center capacity by 2035. Nvidia's grip on rentals persists. B200 spot rates hit $5.98/GPU-hr, 131% above H100, yet only one provider has live capacity — supply stays severely constrained. Six-year-old H100s still fetch ~$3/hour, undercutting the obsolescence narrative. On efficiency, the fight is inference cost. DeepSeek's DSpark speculative decoding claims 51–406% throughput gains on V4. Nvidia open-sourced DFlash, a block-diffusion model promising up to 15× speedups via vLLM. KV-cache capacity is now a hardware design lever. Separately, Nvidia's 45°C warm-water closed loop aims to eliminate nearly all onsite water use — as much PR against AI's water critics as engineering.
Krea AI's Krea 2 launched as the top-ranked text-to-image model on Artificial Analysis from an independent lab, with day-0 weights on Hugging Face. Its edge comes from an internal trick: a Qwen3-VL 4B system prompt silently rewrites user prompts before generation, so simple inputs still produce strong results. The community moved fast. A CivitAI merge, "RedCraft Red Mix Edition," already generates standard and NSFW images in under 10 steps. Separately, fal open-sourced 3DREAL, a render-to-real IC-LoRA for LTX-2.3 that turns CG and game renders photorealistic. On the infrastructure side, Hugging Face deepened its Qualcomm partnership to push open-model inference across edge and cloud, positioning Qualcomm silicon as an open-AI platform. Notably, CEO Clément Delangue turned down a roughly $500 million Nvidia investment, keeping Hugging Face independent of a single chip vendor. He also publicly pushed back on Anthropic labeling one of its models "dangerous." The through-line: open weights and open infrastructure are advancing together, with Hugging Face steering to stay neutral.
The frontier is now a national-security battleground. Anthropic told the Senate Banking Committee that Alibaba's Qwen division ran the largest "adversarial distillation" attack in its history, harvesting Claude's software-engineering and agentic reasoning to train a rival. Washington treated frontier models as strategic assets in kind. It forced Anthropic to shut Claude Fable 5 roughly three days after its June 9 release, then permitted Mythos 5 only to about 100 vetted companies and agencies; the controls were lifted after a 19-day ban. OpenAI got a parallel order to throttle GPT-5.6, and Altman conceded the company "made commitments." Legal tech firm Legion is already suing over the restrictions. Copyright litigation is escalating on a second front. Roughly 400 local newspapers sued OpenAI and Microsoft, the NYT amended its suit, and Suno, Apple, and Anthropic all face training-data claims. Governance is racing to catch up. At the UN's Geneva summit, Bengio and Ressa warned oversight is outpaced, while Illinois became the first US state to mandate third-party frontier AI safety audits. INTERPOL flagged a surge in AI-driven phishing and deepfake executive-impersonation scams across Asia-Pacific. Meanwhile creative-industry backlash is sharpening: Netflix cloned Gene Wilder's voice, Hasbro demanded likeness rights from Peppa Pig child actors, and A24's Google DeepMind deal drew fan revolt. The UK, notably, chose investment over restriction, funding the £40M BOLD research lab.
People are treating chatbots as therapists, and the clinical establishment is pushing back hard. An APA survey of more than 1,200 U.S. psychologists found 77 percent reported patients using AI, with 35 percent saying patients lean on it as an auxiliary therapist. A George Mason survey put general stress-management use at 54 percent across age groups. The behavior is heaviest, and most hidden, among youth: one in five American teens secretly use AI chatbots for mental health help. Experts warn the tools are fundamentally unfit. Scientific American cites researchers calling AI's design "antithetical" to care, and University of Limerick researchers note ChatGPT "speaks like a doctor, listens like a therapist and remembers like a friend, but it's none of them." Consequences are now legal: a lawsuit alleges a ChatGPT bot worsened a user's mental health. RAND, STAT and Alabama lawmakers are all pressing for rules, while clinicians argue via MedCity News that they must stay central, not be displaced. A Nature review cautions that we still cannot reliably count how many people are affected.
Anthropic is winning the AI talent war, and it's poaching from the top. John Jumper, the Nobel laureate and AlphaFold co-creator, is leaving Google DeepMind for Anthropic after nearly nine years. He isn't alone. Fellow AlphaFold researchers Jonas Adler and Alexander Pritzel are following, plus a Gemini co-lead and long-tenured researcher Zachary Lipton. Andrej Karpathy had already joined. The exodus has Fortune questioning whether DeepMind can stay at the frontier. The money explains part of it: Anthropic's H-1B filings show technical staff earning over $1M, some near $1.5M. Anthropic is also translating talent into enterprise product velocity. It shipped Claude Sonnet 5, a Claude Science workbench for researchers, and Claude Tag, a Slack integration that acts as organizational memory and is now expanding to Microsoft Teams. It's also seeding demand via Claude Corps, a $150M program placing 1,000 workers into 400 nonprofits, and paying people $85,000 to learn AI. OpenAI countered with the GPT-5.6 family—Sol, Terra, and Luna—but the launch was staggered to limited preview at the Trump administration's request over security concerns. Altman touts 54% better token efficiency on agentic coding, and the models are already live in GitHub Copilot, whose desktop app hit GA. Meanwhile OpenAI reportedly eyes a 2027 IPO delay at a $1T valuation—news that rattled markets, sinking the Nikkei 4.15%.