Afternoon briefing · 16:00
AI adoption sets the terms at the points of contact
By · bombong
This afternoon's news shows that in AI adoption, the rules governing where models meet the outside world are becoming more important. A retailer built its own entry point and permissions for outside agents, while on the infrastructure side, terms with the power grid and local communities came up together.
Setting rules where AI meets customers
If the morning briefing focused on where models run, the afternoon news is about the rules where those models meet other systems. The clearest example is Walmart. On the 21st local time, Walmart announced it is introducing 'Walmart Agent Gateway' and 'Agent Profile', which let an AI agent of the customer's choice find and buy products from Walmart and Sam's Club 36. In addition to the previously reported purchase permissions and spending limits, it has reportedly also set out how items bought by agents can be returned.
The hard problem in agentic commerce is less the payment technology than the scope of responsibility. Walmart's announcement shows that a seller defining the integration channel, permissions and after-sale rules up front could serve as one reference model.
The same concern shows up at the software distribution stage. Thales launched 'Sentinel Envelope Plus', which makes it harder for AI to reverse-engineer compiled applications or find vulnerabilities in them 31. The company said that in its own tests, an AI agent found 8 of 10 planted vulnerabilities in unprotected software but found none after protection was applied. Because the test conditions and the agent used were not disclosed, this result should be treated as the company's claim.
Infrastructure and local communities become conditions too
Designing these points of contact does not stop at the customer-facing front end; it extends to the infrastructure layer. NVIDIA released European industry adoption cases to coincide with GTC Berlin 35. They include Revolut rebuilding its technology stack around a single foundation model, 'PRAGMA', and public bodies running open models on their own infrastructure, adapted to local languages. NVIDIA said one commerce platform cut inference costs by 90% with open models, but did not disclose the company's name or the baseline for comparison.
What stands out in this announcement is power. NVIDIA claimed that power is currently the biggest bottleneck for AI infrastructure and also presented DSX cases, which design the power grid and the AI factory as one system from substation to rack. This means that where data sits and how power is brought in have become design conditions as important as model choice.
Relations with local communities are also starting to be part of these conditions. According to AI Times, US President Donald Trump posted on the 21st local time that data center operators should reach financial agreements with the communities where their facilities are built 37. However, no plan to formalize this through legislation or an executive order has been reported.
Contact with the physical world is still unverified
Where AI meets the physical world, much remains unconfirmed. The South China Morning Post reported that Unitree Robotics released a promotional video for its 'H2-X' humanoid showing it confronting a knife-wielding intruder and lifting a car to rescue a person 29. However, nothing has been confirmed about whether the scenes reflect real performance, or about specifications and release timing. Using robots for work that involves physical contact with people requires discussing safety standards first.
NVIDIA's addition of training tools, simulation checkpoints and a 'Cosmos Data Factory' blueprint to its Cosmos world model development platform can be seen in the same light 34. It suggests that thoroughly validating robots in virtual environments before sending them into the field has become that much more important.
In the opposite direction, some points of contact are becoming simpler. KittenML released 'Kitten-TTS 2', a 1.58-bit speech synthesis model whose weights take only three values, and a community post said it can do real-time synthesis and voice cloning on CPU alone 30. Quality comparison figures and the license have not yet been confirmed.
Companies adopting AI should design the permissions for outside agents, code protection, power and siting conditions, and physical safety standards first, as carefully as they choose models. Going forward, it is worth watching whether air pollution rules for data centers get tighter, and how the US ban on DeepSeek affects companies' model choices.
This piece was written with the help of AI and includes personal views.
Morning briefing · 08:00
Models spread to local devices, and safeguards follow
By · bombong
Today's news shows that where a model runs has become a central question for both competition and safety. Small speech and coding models quickly reached local hardware and runtimes. A cloud provider presented an agent design that splits work between local and cloud models. On the same day, OpenAI said it would apply the same safeguards to models that reach users through distribution partners.
Models move closer to the hardware
The clearest signal today is how fast small models are reaching user devices and internal servers. On the Hugging Face blog, IBM introduced its speech model Granite Speech 5.0. The company said it handles transcription, translation, and speaker attribution more accurately and faster than before2. A community post said the model is 2B in size. If true, it would suit uses where data can't easily leave the premises, such as meeting transcripts and call centers.
Coding models are going the same way. Ollama v0.40.8 added Recursive's Flint, which runs with a single command5. The release notes describe Flint as an Apache 2.0-licensed MoE (Mixture of Experts, an architecture that uses only some experts for each token) with 262K context. Of its 40B total parameters, only about 2.5B are active per token. Because it reached a leading local runtime one day after its announcement, it is now much easier to try a coding agent on laptop-class hardware.
Cloud providers are building this trend into their designs too. AWS published instructions for building an agent framework called AAIF. Routine tasks go to a local small language model served through Ollama, and only complex requests go to Amazon Bedrock17. In Korea, ESTsoft won a contract from the Jeju Free International City Development Center (JDC). The project puts an sLLM and an agent work platform on the agency's internal servers, isolated from external networks26.
More channels, one set of rules
As models run in more places, providers have started working to make sure the same rules hold on every channel. OpenAI published a post explaining how it keeps safeguards consistent with the distribution partners that offer its models. It said the post is meant to support responsible access and make clear what its Usage Policies expect4. As its models spread to outside clouds, OpenAI is saying the same policies apply on every channel.
Open distribution, where companies bring other firms' models into their own services, is appearing in other markets as well. According to Pandaily, Xiaomi's MiMo-V2.6-Pro is available in Tencent's WorkBuddy agent and on Tencent Cloud TokenHub. The model supports a 1-million-token context and OpenAI- and Anthropic-compatible APIs13. As it becomes common for one company's model to run inside another company's agent, the question of who is responsible for safeguards becomes more important.
OpenAI's cyber defense blueprint, released the same day, takes this question into policy. In the second post of its series on frontier AI governance, the company laid out four pillars1. They are putting frontier AI in defenders' hands, stronger safeguards, modernizing security, and making public institutions more resilient. Developer tools show a similar move. GitHub CLI released expiring, auto-refreshing OAuth tokens as an opt-in public preview19. Where coding agents use gh heavily, this can reduce the risk of long-lived tokens leaking.
What remains unconfirmed
Much remains unconfirmed, though, so it is too early to draw firm conclusions about this trend. Granite Speech 5.0's license and evaluation figures are not in the collected material, and the 2B size comes from a community post. Flint's performance figures come from Recursive's own launch evaluation: 54.9% on Terminal-Bench 2.1 and 74.2% on SWE-bench Verified. There is no record yet of anyone outside the company reproducing these results.
Not all competition is heading local, either. AI Times reported that fal and Black Forest Labs released the video generation model FLUX 3 Max. According to the report, it is 22 times faster than Seedance 2.53. In high-volume content production, how fast and how cheaply a platform can generate may become the main axis of competition. However, the official announcement and evaluation conditions have not been confirmed.
On safeguards, too, the direction is visible but the details are missing. OpenAI's post on distribution partners does not say which rules change for which partners. The cyber defense blueprint also lacks specific policy items in the available material. Whether these principles turn into real contracts and technical controls will have to be checked against future material.
The takeaway for practitioners today is that choosing a model takes more than a benchmark table. You also need to decide where your data lives and which tasks to run locally or in the cloud. You also need to check whether every channel applies the same policies and credential management. Next, watch whether Claude Duet's specifications, so far discussed only in the community, are officially confirmed. It is also worth tracking how the order to remove DeepSeek from U.S. federal agencies affects how companies adopt models.
This piece was written with the help of AI and includes personal views.
Unconfirmed reports 7
Not yet confirmed by an official source. Check the original report.
- OpenBMB's MiniCPM-V 4.7 runs a 35B-A3B MoE on iPhone at 20.5 tokens/s without skipping experts (community report) Original
- Dharma-AI's VocabX specializes small-model vocabularies and claims a 77% cut in production inference cost Original
- Paper 'MemT-1.0' proposes a memory foundation model architecture that builds memory into the model's computation (community report) Original
- GitHub moves the Copilot runtime to Rust with help from AI (report) Original
- "Bank of America hires 1,000 contractors to fix software flaws found by Anthropic's Mythos" (community report, unconfirmed) Original
- "China-linked hackers used AI agents to target 2,250+ cloud tenants; CrowdStrike discloses four more AI hacking campaigns" (community report, unconfirmed) Original
- 'Carbon-X,' a 14.8B open model trained on both DNA and text, released under Apache 2.0 (community report) Original
Community buzz 5
Topics widely discussed on Reddit, Hacker News, GeekNews and elsewhere. These are community reactions and may include unverified claims.
- Claude Duet Claude Duet, Anthropic's proactive personal agent, got 236 points and 204 comments on HN. Posts on Reddit and GeekNews compared it with Meta Muse and OpenAI Dots. The community says it is based on Opus 5.5 and connects to more than 6,000 apps.
- PHP2Shell and AI security Posts kept coming about websites going down one after another because of a PHP remote code execution vulnerability found by AI. One user complained that after patching, OpenAI's cybersecurity restrictions blocked their attempt to run a security check with Codex.
- Local LLM speedups on RTX 3090 On LocalLLaMA, a free trick that reportedly speeds up RTX 3090 prompt processing by more than 20% is getting attention. Other posts included a report of running a 1-bit quantized 748B DeepSeek-V4.1-Flash at 7.2 tokens/s on a €2,000 PC, praise for Qwen 3.8 Flash Next, and news of Flint's release.
- Claude Code for creative and everyday tasks Users posted a stream of reports on using Claude Code and Opus 5.5 for Unity VFX, 3D modeling, real-time 3D airport visualization, game mod performance tuning, and sorting out their parents' TV packages. The author of the TV post stressed that a list of things not to do mattered more than the to-do list.
- Falling AI costs and jobs Several items drew attention together. One post said the cost of top-performing AI fell to 1/20 in six months. A video showed an Anthropic employee telling college students that the jobs they know today won't exist by the time they graduate. Another post revisited remarks about AGI arriving by year-end.