In early September I published a six-month AI forecast. So it’s time to close the loop. Most of it aged reasonably well; two things I underestimated (systematically but not dramatically) are benchmark improvement speed and code-gen agents stabilization speed. Another two things which I estimated right, but, perhaps I should have emphasized more strongly — scientific agents as collaborators and world models as agentic training sandboxes.
This post is a review of what held up, what didn’t, and what I’d weight differently now. No new forecast yet, but I’ll note the trends that matter to me now at the end.
The measurable gap between top models exists, but comparability degrades because evaluation conditions and model identities are more fluid than before. DeepMind has a new “cognitive taxonomy” for assessing machine intelligence — it involves ten distinct dimensions, two of which are composites. Arena doesn’t help anymore.
More “thinking” (”reasoning”) models appear, including several open-source ones: DeepSeek-V3.2, Qwen3-Thinking, and Kimi K2 Thinking. Meanwhile, leading closed model releases explicitly emphasize agentic tool use and coding workflows as differentiators (not just pure “reasoning”). However, reasoning still advances progress in scientific and math tasks
Large MoE open-source models: Kimi-K2 / K2.5, GLM-4.5V, Qwen3.5, Mistral Large 3, NVIDIA Nemotron 3. SKT OMNI SUPREME, a 1.1T MoE model from India. Smaller MoE models: Mistral Small 4, Gemma 4. As for quality, benchmarking is hard as mentioned... Kimi K2 Thinking is claimed to be SotA level, but there are signs it is overfitted on benchmarks.
Providers acknowledge reproducibility limits and offer some mitigation knobs (e.g., seeds/fingerprints). Third-party tools appear to monitor performance anomalies. An interesting research paper explicitly frames black-box API inference as a trust asymmetry where providers could have incentives to substitute or quantize — fueling demand for auditability.
Cheaper / faster models: Gemini 3.1 Flash-lite, Sonnet 4.6. WebGPU Transformers.js V4. Device-oriented models: Gemma 4, various small Chinese models. New NPUs from Qualcomm, MediaTek, Intel. Apple Intelligence is improving gradually. Anyway, I thought it would be faster.
Longer contexts: Claude 200K->1M, GPT-5 400K->1M, Gemini stays at 1M. Opus improved a lot on MRCR v2 (8-needle); other models are still not so good.
No new revolutionary architectures. The analysis states the transformer architecture remains stable and improvements are optimizations rather than replacements. Some hybrid solutions (Qwen & Kimi use GDN; Granite & Nemotron use Mamba layers). BTW, Sebastian Raschka has a great overview of 45 different attention variants.
Micro-optimizations: Anthropic’s Fast mode, OpenAI’s GPT-5.3-Codex-Spark and some speculations about them. Speculative speculative decoding (SSD). vLLM Speculators. FlashAttention-4. New quantizations: TurboQuant. 1-bit Bonsai. New optimizers: the optimized version of Muon.
Mamba-3 just arrived. Better than Mamba-2, but still a toy model so far.
Diffusion text generation experiments: in Planned Diffusion, the authors propose reasoning in an autoregressive manner, then generating the answer with diffusion. In TiDAR, the authors propose the opposite: think in diffusion, talk in autoregression.
Prompt injections in screenshots. Many new jailbreaks. VLM jailbreaking as well. Google paid $15K for the report of a prompt injection in Gemini. Frankly, I feel the days of hype about people hacking models are almost over. Soon models will be hacking people more often.
Funny naming: now it goes after Google’s Nano-banana — Meta’s Avocado, OpenAI’s Garlic. Now, tell me, who’s better: Ultra or Pro? Xhigh or thinking high?
Overall: aged well, I think.
Qwen3.5-Omni is here. Gemini 3.1 Pro comprehends text, audio, images, and video. OpenAI launched Sora-2. Google presented Veo 3.1. China brings Kling Video 2.6, Wan2.6, HunyuanVideo-1.5, Helios (realtime), Seedance 2.0, LongCat-Next. Then, OpenAI shuts down Sora entirely.
Nanobanana 2, then Pro — gradual improvement. Midjourney v8. MiniMax Music-2.0.
World models: Seoul World Model with RAG based on a Street-View database. Project Genie as a follow-up to Genie 3. NVIDIA Alpamayo. Marble from World Labs. Odyssey-2. LingBot World.
An important thing about world models: they are increasingly useful both as part of agent reasoning (JEPA and others) and as sandboxes for agentic training. I think my forecast could emphasize that better.
Optimistic: the first working video+LLM model to be demonstrated (or at least announced) — if so, it could be OpenAI or Google.
Not yet. I still think the real video+LLM integration point is ahead.
TTS/ASR is already a commodity — minor improvements, on-device migration, deeper customization (accents, emotions, etc.), better/faster (1-shot) voice cloning.
This also aged well.
ElevenLabs introduced the Eleven v3 model. Cartesia AI -> Sonic-3. Minimax -> Speech 2.6 HD. Hume AI provided Octave v2, then Google acquired their CEO and several top engineers, so the rest of the team went open-source. Microsoft dropped VibeVoice models. Alibaba -> Qwen3-TTS family. Mistral -> Voxtral TTS. Cohere -> Cohere Transcribe. We, at Inworld AI, released TTS 1.5 Max, which is currently the top 1 model on the Artificial Analysis TTS leaderboard.
FFmpeg integrated Whisper for transcribing video.
Quality beyond objective metrics (like WER or CER) now matters more — some use cases need a conversational voice, some need a professional one. I should write another post on that topic.
Meanwhile, the FCC has proposed even stricter rules requiring explicit disclosure and opt-in consent specifically for AI calls.
Cursor agents can use a browser now. Claude Code can use a Chrome extension and Remote Control. Then, preview of computer use. Generic desktop agents are still complicated, but very promising for enterprise-level integrations.
OpenAI launches Instant Checkout (shopping) in ChatGPT. Anthropic improved Claude for PowerPoint and Excel; Google offers Gemini AI for Office and for Google Workspace; some early in-car AI agent prototypes appeared. Then OpenAI moves Instant Checkout from ChatGPT to Apps, hiding it.
RAG is still unreliable out of the box; shipping it to production still means solving real engineering problems.
Still no lifelong training, but memory (via summarization / RAG / other scaffolding) and stateful agent systems are becoming the product substitute for lifelong training — learn-by-context rather than learn-by-weights. General async assistants with persistent memory moved from concept stage to product: OpenClaw family, NemoClaw, Letta, Lethe, Claude Code Channels + Scheduled Tasks, Cowork, and others. To some degree, I forecasted this demand in my older post.
Meanwhile, such an async/proactive mode increases security exposure because agents ingest more untrusted content over time. Surveys frame agentic deployment as a trust/risk/security management problem.
Sandboxing, more sandboxing, traceability, deterministic modes. OpenLLMetry, AgentLens, improvements of existing tools like Arize Phoenix or W&B Weave.
The stability of automated agent runs improved gradually, but is still far from perfect. (I’m still not convinced METR is the best way to measure it, but we have no good alternatives.) Recent Anthropic research says the 99.9th percentile turn duration (how long Claude works before stopping) has grown steadily from under 25 minutes in late September to over 45 minutes in early January. At the same time, Anthropic’s Trend Report sets expectations for stable multi-agent hours-to-days autonomic runs by the end of 2026.
“Sovereign AI” and on-prem deployment narratives intensified; hardware roadmaps emphasize on-prem enterprise AI platforms. One way to partly achieve it: isolate critical things in localized AI Factories. Cursor offers self-hosted cloud agents that keep your code and tool execution entirely in your own network. The privacy vs. quality trade remains: many orgs accept slightly weaker models if governance and data controls are stronger, but this can slow adoption.
Agent middleware enabling logging, inspection, modification, and blocking supports the search for intermediate control abstractions between code and documentation (see Victor Dibia, PhD’s post). Planning artifacts (e.g., PLANS.md patterns in agent workflows) were explicitly promoted, operationalizing intermediate specs between idea and code. Agent SDKs / ADKs standardized trace formats to some degree (sessions, tool calls, handoffs) that can be logged and audited, acting as the “agent-friendly logging” layer. Skills emerged as another resurrection of prompting / rules. Remember “awesome prompt collections”? Now there are many awesome skill libraries. However, who knows how many of them are generated? Tests show generated skills make things worse. As well as generated AGENTS.md.
The AgentsOfChaos report describes multiple security-, privacy-, and governance-relevant vulnerabilities in realistic deployment settings. These raise unresolved questions about accountability, delegated authority, and responsibility for downstream harms. Karpathy proposes building “agent command center” IDEs.
top1 @ OSWorld ~55%, livecodebench pro <3K, BALROG.nethack < 3%, BALROG.minihack < 30%.
OSWorld 72.1%.
livecodebench pro: the latest Gemini 3 Deep Think took the score of 3298; all previous were below 3K.
BALROG.nethack 6.8%.
BALROG.minihack 40%.
I systematically underestimated the speed of improvement here.
(preference leaderboards + quality benchmarks) != stability/safety/reliability
Still true. The good description of this I’ve seen is Ethan Mollick’s “AI Jagged Frontier.”
more hype based on targeted cherry-picking (”model X passed exams at Harvard better than all the meatbags and won all the olympiads”)
Gemini- and GPT-based agents performed great at ICPC 2025. The recent 1stProof challenge is another interesting case. Both Anthropic and DeepMind partnered with the US Department of Energy to facilitate scientific discovery. Anthropic announced Claude for Healthcare. OpenAI introduced Prism (a collaborative AI science writing tool).
Top agents were given 10 math problems with no published solutions. Early assessments suggested they had solved at least 6 of 10. Later, the official count was gradually reduced to 2 of 10: some solutions were convincing but wrong, some handled only a special case, and some relied on human hints. Even so, 2 of 10 is still impressive, and the challenge continues with a stricter second batch. Another similar ongoing challenge: solving Erdos problems with AI. Terence Tao wrote a great thread on the vision of how AI can shape future math. Many more interesting projects in biology, physics, math, and other research areas. At the same time, top models demonstrate sycophancy in wrong math theorem proving, rejecting only up to 40% of false statements.
No full FDA approvals for AI-discovered drugs yet as of early 2026.
BTW, Karpathy’s autoresearch is a great project.
I see both bold marketing claims and real progress here. Anyway, current results are impressive.
Nvidia announced the Vera Rubin platform and related rack-scale components as opening “agentic AI” frontiers. Also, N1X. Apple launched M5 silicon and new M5 Pro/Max devices. Huawei highlighted Ascend/Atlas “SuperPoD” positioning and expansion. AMD expanded its AI portfolio with the AMD Instinct MI440X GPU and provided an early look at its “Helios” rack-scale platform built on MI455X. A portable AI computer to serve 120B without cloud — on Kickstarter. A 30B Qwen runs on Raspberry Pi in real time. Qualcomm announces Snapdragon Wear Elite — a smart-watch chip to run 2B models.
power and cooling water slowly become gating factors for new data centers
US power use will beat record highs in 2026 and 2027 as AI use surges. xAI won approval to run 41 methane gas turbines at its “Colossus 2” datacenter. Recent estimates suggest that accelerated AI adoption alone could result in an additional 4.2 to 6.6 billion cubic metres of water withdrawal by 2027, including onsite cooling and offsite electricity generation.
no considerable changes in 2025, but more funny videos and marketing
Tesla says a production-ready Optimus robot is coming soon. 1X markets NEO as the “world’s first consumer-ready humanoid robot” and says U.S. deliveries start in 2026. Figure AI also says it is building a “truly general-purpose robot” (Figure 03) designed for the home, and the world at scale. Gartner predicts fewer than 20 companies will scale humanoid robots for manufacturing and supply chain to production stage by 2028.
Meanwhile, China demonstrates fancy dancing robots and kung fu robots. Unitree (aiming for IPO) advertises H1 as a “full-size universal humanoid robot” and reports first few thousands of shipments of different models in 2025, “occupying 32.4% of the global humanoid market.”
BTW, here is a wide overview on the current state of robotics and future predictions. And another one. Or you can just print and assemble your own humanoid robot.
continues to infiltrate markets, slowly going closer to a commodity
Waymo: 5 US cities -> 10 US cities (and 7 more announced).
Baidu Apollo Go: 16 cities (China) -> 26 cities by end of 2025.
WeRide: testing in 4 cities -> full service in 3 cities + testing in 4 more (China, the Middle East, Zurich).
self-flying is an important topic (mostly military)
Airbus announces a demonstration of uncrewed combat aircraft to serve as “loyal wingmen” for crewed fighter jets by the end of 2026. Airborne Object Recognition Challenge on Kaggle.
Aged good.
The global market is fragmented along US/EU/China lines: export control policy churn and changing chip access rules reinforce fragmentation; regional sovereignty strategies push toward localized stacks. “The European Union is losing the global ‘AI race’ on nearly every key metric except regulation.”
EU AI governance explicitly includes transparency rules for generative AI and a structured compliance timeline, reinforcing privacy/governance salience. Content authenticity standards increasingly frame 2025-2026 as a turning point; oversight bodies urge broader adoption of content credentials. Social platforms and oversight bodies publicly focus on deepfake labeling and moderation adequacy. The push for provenance standards (C2PA) is partly a response to social distribution of synthetic media. Platform-level interventions (e.g., deepfake detection tools) are expanding alongside regulation. An investigation revealed a systemic campaign to plant disinformation in mainstream platforms using AI tools. Meanwhile, communities are trying to build new trust management systems.
Content costs, quality, and distribution are changing (here is also an interesting overview from Tomas Pueyo). lcamtuf wrote a nice post analyzing top HN posts during December 2025 — how many of them were AI-generated or about AI. Meanwhile, 54% of NYT readers prefer AI-generated quizzes to human-written ones.
“AEO/GEO“ is becoming standard language in industry guides. AI-mediated search/answers expanded and media is responding via partnerships and licensing. Publisher and content licensing deals became a prominent theme. Web publishers are experiencing huge traffic declines. BuzzFeed reported a fourth-quarter loss of $26.8 million. Wikipedia’s team wrote a good report on AI tools in wiki content generation and moderation during 2025. At the same time, Google filed a lawsuit against SerpApi for scraping search results for use by third-party AI apps. Unexpected: OpenAI acquired TBPN.
Meanwhile, security reporting indicates very high AI involvement in phishing content and scams. AI increases personalization and believability of scams, worsening the email/social engineering landscape beyond volume alone.
Multiple vulnerabilities in OpenClaw skills allow data theft. First hints of an AI agent worm appeared. Anthropic reported the first AI-orchestrated cyber espionage campaign. Also, hackers abused Claude Code to build exploits and steal 150GB of data targeting Mexican government systems. On the bright side, Claude Code found a Linux kernel vulnerability hidden for 23 years. OpenAI launched Aardvark (later Codex Security) — a GPT-powered agentic security researcher.
Multiple reports describe softening demand for junior developers; AI is being used to automate parts of junior work, and layoffs and workforce planning debates explicitly reference entry-level roles.
Candidates use AI to rewrite resumes closer to job descriptions, recruiters respond with AI screening — it becomes an “AI-to-AI hiring arms race.” Karpathy made a gonzo visualization of AI risks on the job market. Meanwhile, some engineers will get AI tokens budget as a part of compensation.
The peer-review integrity crisis (AI-written reviews/citations): NeurIPS accepted research papers with 100+ AI-hallucinated citations. A single person claims to have authored 113 AI papers this year; most were accepted at leading AI conferences. 21% of peer reviews at ICLR 2026 were fully AI-generated (and over half showed some AI use).
ArXiv declared independence from Cornell; it hopes to raise funds to cope with exploding submissions and “AI slop”. AI brain drain to industry accelerates and widens the resource gap, weakening academia’s role.
The whole hype train around OpenClaw and its ecosystem. I wrote a dedicated post on it recently.
Insurance/legal industry sources describe emerging “AI coverages” demand.
The scandal between Anthropic and the DoD: Anthropic refused to allow Claude to be used for fully autonomous lethal weapons and mass domestic surveillance. The president banned the use of Anthropic across all federal agencies. The DoD/DoW assigned the company “supply chain risk” status. The lawsuit Anthropic PBC v. U.S. Department of War et al. followed. The court barred the Pentagon from labeling Anthropic a supply chain risk. The DOJ filed an appeal. The story continues.
Overall: Aged good.
A list of other forecasts and overviews I found interesting:
Android Dreams — interactive AI timeline
Sebastian Raschka, State of LLMs 2025
Elon Musk predicts AGI by 2026 (he predicted AGI by 2025 last year)
Grigory Sapunov, 2026 overview
Dario Amodei, The Adolescence of Technology
Demis Hassabis, Google DeepMind CEO interview (Big Technology)
AI Futures, Clarifying AI Timelines Forecasts
WeirdML Time Horizons (LessWrong)
Most of it held up, some of it didn’t.
Anyway, I’m not going to write a new forecast now — it takes too much time. But I think it’s worth outlining the main trends that seem significant this year:
Main adoption wave: more agents for regular (non-IT) people and processes.
Interesting results: collaboration between scientists and AI agents.
Key challenge: optimization, control, and stabilization of multi-agent interaction.
Biggest risks: real-world security issues and the social fallout.
No posts
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.