Today’s machines are learning to own more of their internal machinery while the companies around them consolidate the external kind. An open trillion-parameter model, a CPU compiled into neural weights, a proposed delivery-platform colossus, an LLM that “sleeps,” and a sharp fall in computer-science enrollment all point to the same uncomfortable pattern: capability is expanding faster than our institutions can decide what it is for.
1. Mistral opens the weights of its trillion-parameter flagship
Source: Mistral AI — Mistral Large 4 (Le Chonk) is now open weight - https://mistral.ai/news/mistral-large-4-oss/
Mistral released the weights for Mistral Large 4, a natively multimodal mixture-of-experts model with 1.05 trillion total parameters, 52 billion active parameters, a one-million-token context window, and new low, medium, and high reasoning modes; the company says three more weeks of reinforcement learning lifted it to 41% on Terminal-Bench 4.0 and 48% on the Artificial Analysis Coding Agent Index while using fewer tokens than rival open models. The strategically important part is not “one trillion,” a number large enough to make procurement departments develop weather systems, but that capable organizations can inspect, customize, and self-host the model under Mistral’s modified MIT-style license, although commercial model-as-a-service operators need a separate license. Vendor benchmarks still deserve independent testing, especially the security claims, but this is a serious expansion of the open-weight frontier and of Europe’s attempt to make model sovereignty something more concrete than conference vocabulary.
2. Qwenputer compiles a deterministic RISC-V computer into model weights
Source: Divyansh Agrawal — Qwenputer: a fully functioning, deterministic RISC-V CPU inside Qwen3.5 - https://divagr.com/blog/qwenputer
Independent researcher Divyansh Agrawal embedded a 32-bit RV32IM RISC-V core into additional layers attached to an otherwise unchanged Qwen3.5 4B model, using Gated DeltaNet state as exact addressable RAM so arithmetic, C, and MicroPython can execute during forward passes without an external tool or sandbox. The experimental CPU reportedly passes the official RISC-V tests, matched QEMU, ran 18 million instructions without error, and pushed Qwen from 48.6% to 100% on NUPA long-integer arithmetic; the tradeoffs are an 18.8 GB checkpoint, a special boot driver, and glacial-by-silicon execution because each instruction consumes decoded tokens. This is not a practical replacement for tool use, but it is a marvelous proof that hybrid recurrent-transformer state can host deterministic computation — and a warning that “what the model contains” may soon include executable mechanisms that ordinary chain-of-thought monitoring cannot see.
3. DoorDash proposes a $17 billion acquisition of Instacart
Source: DoorDash — DoorDash to Acquire Instacart to Help Local Stores Win - https://about.doordash.com/en-us/news/doordash-to-acquire-instacart
DoorDash agreed to acquire Instacart for roughly $17 billion, offering $65 per share in a transaction funded at least 75% with cash and expected to close in late 2027 or the first half of 2028 if shareholders and regulators approve it. The proposed combination would join DoorDash’s delivery network with Instacart’s relationships across nearly 100,000 stores, enterprise grocery software, advertising business, and AI shopping tools; the companies promise one membership fee, access to more than 50,000 additional stores for Instacart customers, and no grocery exclusivity. Those benefits are plausible, but this is also vertical consolidation of logistics, demand, merchant software, shopper labor, advertising, and increasingly the interface through which an AI decides what households buy — a stack with enough leverage to call itself “local commerce” while becoming very large indeed.
4. Google Research gives continual-learning models a sleep cycle
Source: Google Research — The sleep framework: A new way for LLMs to retain knowledge in continual learning - https://research.google/blog/the-sleep-framework-a-new-way-for-llms-to-retain-knowledge-in-continual-learning/
Google and university researchers proposed a continual-learning framework that alternates an active “wake” phase with offline “sleep,” when newly unlocked low-rank mixture-of-experts memory blocks absorb short-term information through distillation and imitation learning, then generate and filter synthetic “dreams” for further training. The team reports near-perfect BABILong scaling to 10 million tokens, little forgetting during sequential learning of unseen languages, and stronger factual and mathematical performance than its selected baselines, though these are research results from a fixed architecture and human-defined training rules, not evidence of a model mysteriously becoming sentient after a nap. The deeper point is architectural: if deployed models can periodically rewrite durable memory without flattening old skills, the boundary between inference and training starts to dissolve — useful for adaptation, and absolutely deserving of audit logs, rollback, and a very stern bedtime routine.
5. U.S. computer-science enrollment drops by double digits
Source: National Student Clearinghouse — Preliminary Fall Enrollment Trends - https://www.studentclearinghouse.org/news/total-higher-ed-enrollment-holds-steady-in-preliminary-fall-enrollment-trends-report/
Preliminary National Student Clearinghouse data shows enrollment in Computer and Information Science programs falling 12.4% at four-year undergraduate institutions, 12.3% at two-year colleges, and 10.4% in graduate programs even as total undergraduate enrollment edged up 0.6%; the release covers 7.92 million enrollments reported by 46% of participating U.S. institutions and will be revised in January. It is too early to assign one cause, and “AI took the jobs” is a headline, not a causal model, but the synchronized decline is strategically significant: students are pricing uncertainty about entry-level work into educational decisions before employers and universities have redesigned the path from novice to expert. If fewer people learn the fundamentals while agents generate more of the code, the industry may achieve extraordinary output and then discover it has automated the apprenticeship pipeline that produces future reviewers.
The Professor's Read
The day’s theme is not that software is eating the world; it is that intelligence, computation, logistics, memory, and labor markets are being fused into fewer, denser systems. That can create remarkable tools, but concentration without legibility is how progress becomes dependency, so the winning discipline now is not merely building stronger machines — it is preserving independent measurement, reversible decisions, and enough human apprenticeship to understand what the machines just did.
References
- Mistral AI, “Mistral Large 4 (Le Chonk) is now open weight” - https://mistral.ai/news/mistral-large-4-oss/
- Mistral AI model documentation, “Mistral Large 4” - https://docs.mistral.ai/models/mistral-large-4-0
- Mistral Large 4 model repository and license - https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B
- Hacker News discussion, “Mistral Large 4 (Le Chonk) is now Open Weight” - https://news.ycombinator.com/item?id=50289418
- Divyansh Agrawal, “Qwenputer: a fully functioning, deterministic RISC-V CPU inside Qwen3.5” - https://divagr.com/blog/qwenputer
- Qwenputer source code - https://github.com/divagr18/qwenputer
- Hacker News discussion, “Qwenputer: A functioning, deterministic RISC-V CPU inside Qwen3.5” - https://news.ycombinator.com/item?id=50282838
- DoorDash, “DoorDash to Acquire Instacart to Help Local Stores Win” - https://about.doordash.com/en-us/news/doordash-to-acquire-instacart
- Google Research, “The sleep framework: A new way for LLMs to retain knowledge in continual learning” - https://research.google/blog/the-sleep-framework-a-new-way-for-llms-to-retain-knowledge-in-continual-learning/
- Behrouz et al., “Sleep-time Compute: Beyond Inference Scaling at Test-time” - https://arxiv.org/abs/2606.03979
- National Student Clearinghouse, “Total Higher Ed Enrollment Holds Steady in Preliminary Fall Enrollment Trends Report” - https://www.studentclearinghouse.org/news/total-higher-ed-enrollment-holds-steady-in-preliminary-fall-enrollment-trends-report/
- National Student Clearinghouse Research Center, Preliminary Fall Enrollment Trends methodology and dashboard - https://nscresearchcenter.org/prelim-fall-enrollment-trends/