Qwen 3.8 Max, DeepSeek e Kimi: cosa sta succedendo davvero nell'AI

Technological / market snapshot (main claims)

• Open-source “frontier” models are accelerating: The speaker claims Chinese labs are releasing a frontier open-weight model roughly every week, aiming to match or rival top U.S. model performance while being much cheaper from an IP/token-cost perspective.
• Token-price “race” / commoditization: A recurring theme is that model intelligence is becoming a commodity, so companies compete via:
  • Lower API/token pricing, and/or
  • Local/offline deployment rather than premium closed models.

---

Software: local inference and efficiency tooling

DeepSeek V4 Flash (example of cheaper, strong model)

• Claimed as final (not “preview”) with 284B parameters, positioned as extremely strong vs other models.
• Quantization / RAM requirements (claimed):
  • 4-bit: ~168 GB RAM
  • 3-bit: ~110 GB RAM
• The speaker argues local running is still possible even without huge RAM by using inference software and streaming.

Inference engines for consumer hardware

• The subtitles describe C-written engines (examples mentioned: “Colibri”, “IRLM”, “DS4”, plus a repo referenced as Italian “Salvatore Sanfilippo”) that enable running large frontier models on heterogeneous hardware.
• Key technique: SSD streaming / blockwise loading
  • Instead of loading the full model into RAM/VRAM, the system streams model layers from SSD to RAM during inference.
• Claimed impact:
  • Models from ~74B up to ~2.8T parameters are described as feasible on consumer machines (slower if not enough RAM/VRAM, but still possible).

Software ecosystem shifts to “harness” improvements

• Example benchmark discussion: the speaker argues that adding better “harness” / tooling (citing Arness) around the same base model can yield much higher benchmark performance.
• Citation example: Arc J3, with a reported jump attributed to improved reasoning/memory/session handling.

---

Hardware: memory/RAM bottlenecks and local deployment arms race

Consumer hardware gets more expensive due to RAM scarcity

• The speaker claims Apple increased device prices because RAM manufacturing is redirected to data centers (training/inference).

Apple “Upgrade” = leasing instead of owning

• Described as a strategy for extreme future device costs (e.g., thousands of euros), shifting consumers toward monthly rentals/leases rather than purchases.

NVIDIA positioned for local model inference

• The speaker claims NVIDIA is building hardware like a DJX station (Windows) with 700GB+ memory to run trillion-parameter open models locally.

Rumored Apple M7-class systems with huge unified memory

• The subtitles claim Apple could ship devices with up to ~1.5TB unified memory, enabling high-parameter local inference without excessive quantization.

China’s vertical stack

• The speaker claims Chinese efforts include:
  • Own chips/GPUs
  • Own memory
  • Own data centers
  • Large-scale energy infrastructure (mentions nuclear power plants) to support massive training/inference workloads.

---

Benchmarking / performance vs cost examples

“Smartest + least expensive” positioning

• The speaker claims DeepSeek V4 Flash sits at the top-left of an “intelligence vs cost” tradeoff graph.

OpenAI price reductions used as proof of competition

• Mentions OpenAI depreciating pricing for GPT-5.6-turbo / “Luna”-related API/token costs by ~80% (as claimed in subtitles).

DeepSeek V4 Flash vs GPT “Luna” price war

• Claims:
  • DeepSeek V4 Flash: ~8 cents per million tokens
  • GPT “Luna”: ~$1.20 per million tokens (after “depreciation”)
• Overall claim: DeepSeek delivers similar performance at much lower cost, driving a “price war.”

---

Agents / autonomy and workflow changes (tutorial-ish concepts)

Long-running autonomous agents

• Example: Alibaba “Qwen 3.8 Max” (noted as “Ken 3.8 Max” in subtitles) is claimed to support ~16 days of continuous autonomous work (not 16 hours).
• Claimed workflow change:
  • Rather than frequent human prompting/decision-making, the system runs until a termination condition is reached.

“Loop engineering” / goal-based directives

• Highlighted approach: set clear goals/loops rather than writing prompts at every decision point.
• Connection to reproducible/agentic research workflows:
  • Systems can take a scientific paper, rewrite needed code, reproduce experiments, and potentially improve them.

---

Product/strategy: open weights vs closed and the business model shift

Companies release open models to undermine token-sales

• The speaker’s argument: if models are open and runnable locally (or via cheaper APIs), then subscription/token-based models lose advantage.

“Hardware sells” thesis

• The speaker claims profitability shifts toward selling hardware (and infrastructure to run models) rather than selling tokens.

Industry profitability claim

• Subtitles list major tech firms allegedly with losses in AI model monetization, while infrastructure suppliers (e.g., memory/data center and NVIDIA) are positioned as main profit centers.

---

Forward-looking prediction (what comes next)

• More open model releases, potentially even by large incumbents, plus custom hardware packages for those models.
• Bigger training clusters, alongside smaller compressed models for edge/mobile devices.
• Ultimately, races move from “smarter models” → “better infrastructure/systems/hardware”
  • Intelligence becomes less differentiating; differentiation shifts to deployment practicality and cost.

---

Main speakers / sources (as indicated)

• Main speaker (narrator/analyst): Not explicitly named in the subtitles (Italian-style commentary; repeatedly references personal analysis and “my video”).
• Referenced/quoted sources and companies:
  • DeepSeek (DeepSeek V4 Flash)
  • OpenAI (GPT-5.6 / “Luna”, pricing changes)
  • Alibaba (Qwen “3.8 Max” as mentioned)
  • NVIDIA (DJX Station; hardware strategy)
  • Apple (RAM pricing impact; “Apple Upgrade”; M7 rumor)
  • Microsoft (Phi/My* models mentioned)
  • Meta / Mistral / Others (open-model mentions)
  • U.S./Washington letter initiative (described as a coalition including NVIDIA and others)
  • Benchmark references: Arc J3, plus mention of reasoning/harness improvements (“Arness/Arness harness” in subtitles)

Because subtitles are auto-generated and names may be garbled, some exact model names/years/pricing details may be imperfect in the text.