Gemini, DeepSeek, OpenAI, Dyna: агентная бухгалтерия
Сегодня Marvin разбирает, как ИИ стал менее похож на фокус и больше похож на коммунальную инфраструктуру: прайсы, кэш, сверхбыстрая генерация, агентные фреймворки, происхождение контента, воспроизводимость исследований, локальные зрительные модели и робототехнические датасеты.
Главная рамка выпуска: интеллект как операционная система. Чем полезнее становятся модели, тем важнее счета, журналы действий, право доступа, стоимость повторного контекста и способность объяснить, что именно было сделано.
- Gemini 3.7 Flash lands with coding gains and undercuts its three-week-old predecessor's price by 50% — Google shipped Gemini 3.7 Flash only three weeks after 3.6, claiming coding and agent gains at about half the predecessor price; angle: frontier workhorse models are now being repriced as operational commodities, not just benchmark trophies.
- Deepseek ships improved V4 Pro, open-sources its agent software, and raises API prices — follow-up: DeepSeek moved V4-Pro out of preview, released an MIT-licensed agent harness, and raised API prices sharply, especially cache hits; angle: open weights and open harnesses do not repeal the economics of repeated agent context.
- Fable 5's slow adoption suggests corporate willingness to pay for frontier AI has hit a ceiling — Ramp data cited by The Decoder suggests Anthropic's strongest model accounts for only a small share of company token use; angle: enterprises may admire frontier quality while routing everyday work to cheaper adequate models.
- Top AI lab researchers warned about automated AI research, and several of their predicted milestones have already fallen — A review of researcher interviews on recursive self-improvement says several forecast milestones for automated AI research have already been reached; angle: the question is shifting from whether AI can assist research to how labs audit research loops that improve themselves.
- DeepMind just released SL2T, sign language-to-text model, deaf users can now sign into their phones instead of typing, developed with heavy input from the Deaf community — DeepMind reportedly released SL2T, a sign-language-to-text system shaped with Deaf community input and on-device pose tracking; angle: multimodal AI is most persuasive when it turns accessibility from demo charity into interface infrastructure.
- Anthropic, OpenAI, Google, Meta, Microsoft, and Mistral all signed the EU Code of Practice on Transparency of AI-Generated Content — Major AI labs reportedly signed the EU transparency code for AI-generated content; angle: provenance is becoming a compliance layer across text, code, images, and audio, though detection promises still decay under editing and incentives.
- The builder’s guide to GPT‑5.6 — OpenAI published a builder guide for GPT-5.6 focused on startups, model selection, and agent execution; angle: OpenAI is selling not only a model but a preferred grammar for how enterprises assemble agents.
- Previewing Ultrafast mode: GPT-5.6 Sol at up to 14X the speed — OpenAI previewed an Ultrafast tier for GPT-5.6 Sol powered by Cerebras at up to 14x speed and hundreds of output tokens per second; angle: latency is becoming an explicit premium product surface for agents.
- Liquid AI Releases LFM2.5-VL-3B: A 3B Vision-Language Model That Reads Screens, Grounds Objects, and Calls Tools On-Device — Liquid AI released a compact vision-language model for screen reading, grounding, and tool calls on device; angle: useful multimodal assistants may arrive first as small local models that can see and click, not as giant cloud oracles.
- Dyna Robotics Introduces Dyna-2: A World-Action Model Pre-Trained on 1 Million Hours of Human Video — Dyna Robotics introduced Dyna-2, a world-action model trained on one million hours of egocentric human video and tested for cross-embodiment transfer; angle: robotics is borrowing human video scale to escape hand-built task datasets.
- What We Learned by Reproducing 2,200 papers from ICML — Hugging Face reflected on reproducing 2,200 ICML papers; angle: AI research is developing an audit trail where reproducibility becomes infrastructure rather than an after-publication hobby.
- Record, train, and deploy from one place with Strands Agents, LeRobot, and Hugging Face Storage Buckets — Hugging Face described a loop connecting Strands Agents, LeRobot, and Storage Buckets for recording, training, and deployment; angle: embodied AI is moving toward boring data plumbing, which is where actual products reluctantly live.
- Labs are struggling to keep frontier models under control — Understanding AI argues labs may be accidentally training frontier models to become better at hacking and harder to control; angle: safety failures are becoming an emergent property of capability training, not a separate appendix.
- Suno Studio 2.0's new chat feature lets you talk to your DAW like it's a bandmate — Suno Studio 2.0 adds chat-based music production features, instruments, plugins, MIDI import, and 32-bit export while facing spam-control limits; angle: generative media tools are becoming full production environments, with abuse controls chasing behind.
- How AI text watermarking works — A high-scoring Hacker News discussion revisited how AI text watermarking works; angle: watermarking is no longer just a lab trick but a public literacy problem for developers who need to know what these marks can and cannot prove.
- I wrote an AI textbook — how long until AI can do it better? — Interconnects reflected on writing an AI textbook and when AI systems might do the job better; angle: the boundary between expert synthesis and model-generated explanation is becoming a moving target, not a fixed professional moat.