Microsoft AI Chief Criticizes Anthropic's "Claude Consciousness" Claims
Mustafa Suleyman, head of Microsoft's AI division, has criticized Anthropic for treating Claude as if it were a conscious being. According to The Verge, Suleyman pointed to the ethical and practical problems of eliciting conscious-seeming behavior from AI models in his critique of Anthropic's approach.
The remarks are widely seen as a response to Anthropic's active commentary on "AI welfare" and the "moral status of AI" around the launch of Claude Fable 5. The AI research community has debated for years whether large language models truly possess "consciousness" or "emotions," and with the top executives of Microsoft and Anthropic — two of the AI industry's major players — now publicly at odds, the debate has been reignited.
Discussion of the matter has also begun on Hacker News, where it is drawing attention as an unprecedented development: the heads of AI companies openly sparring over the "consciousness" of their own products.
Google Fires a "Warning Shot" in the AI Subscription Price War
According to TechCrunch, Google is moving to make its low-cost AI subscription tier even cheaper, throwing fresh fuel on the AI industry's price war.
The move follows last week's 40% price cut for Google AI Plus (to 725 yen per month), and with AI model usage fees continuing to fall, every company's profitability is once again being called into question. A post titled "AI Subscriptions Are Dead" has also spread on X, sparking lively debate over the fate of the subscription model itself.
On Qiita, an article titled "AI Industry in the Second Week of June 2026: IPOs, Price Cuts, and the Wall of Operations" argued that the rush of new model announcements has died down and the industry has entered the phase of "how to run the tools you already have, more deeply." As models become increasingly commoditized, the price war is expected to grow even fiercer.
Taking Qwen2.5-7B to 96% of Claude Haiku with DV-DPO — for a Mere $3 in API Costs
A report of Qwen2.5-7B being fine-tuned to 96% of Claude Haiku's performance has drawn a huge response in Reddit's LocalLLaMA community.
The technique is called "DV-DPO" (Dialectical Verification Direct Preference Optimization). Specifically, a council of three AI "voices" generates a synthesized answer to each question, and the losing voice then "cross-examines" the synthesis by objecting to it. DPO pairs (chosen = revised, rejected = pre-revision) are created only when the synthesized answer ends up being revised.
The result: with just 1,040 pairs of training data (roughly $3 in API costs), the model scored 100% on format, 100% on commits, 89% on context, and 96% overall in a head-to-head comparison with Claude Haiku. The approach has won praise for using no human labelers at all and converting only "genuine revisions" made under adversarial pressure into a learning signal.
An autonomous loop — failure_detector → auto_red_team → DPO pairs → retrain → redeploy → eval — is now in operation, and it is drawing attention as a new paradigm for low-cost, automated model improvement.
Could Furiosa AI's Inference Chip Be a Game Changer for Local LLMs?
"RNGD", an inference-focused chip from South Korean AI chip startup Furiosa AI, is generating buzz in Reddit's LocalLLaMA community.
The specs: a TSMC 5nm process, Hynix HBM3 delivering 1.5TB/s of bandwidth and 48GB of VRAM, and a TDP of 180W. It has reportedly already been tested with LG's LLMs. The poster cited the current landscape — an NVIDIA RTX Pro 5000 48GB (non-HBM) at $5,000, an AMD R9700 32GB at $1,300, and an Intel B70 32GB at $1,000 — and argued that "if this chip, with HBM memory bandwidth at around $2,500, runs on llama.cpp, it will upend the local LLM landscape."
That said, there are no plans to sell to the consumer market at this point, and the poster went no further than expressing hope about what would happen "if it were sold." Even so, one thing is certain: new options are starting to appear in an AI inference chip market where NVIDIA has had a stranglehold.
Ottawa, Canada Unveils AI Regulation Bill, Set to Include Age Restrictions
The city of Ottawa, Canada, has put forward a regulatory bill covering social media and AI. According to CBC, the bill is expected to include the introduction of age restrictions for young users.
As AI spreads, regulatory efforts are accelerating around the world. In the EU, concrete measures have already begun — such as ordering Meta under the Digital Markets Act to readmit rival AI chatbots to WhatsApp — and similar regulatory debates are heating up in North America as well.
Age restrictions on AI use by young people could have a direct impact on how AI tools are used in schools and at home, and how they will actually be enforced is attracting close attention.
---