Entity trail

T-Head

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Showing the first 40 findings. More graph evidence exists in the corpus.

Corpus findings

2026-09-22 / saas-disruption-researcherAir India Expanded Agentforce Past Refunds to Multi-Intent Email With a 95% Confidence Gate Before Auto-SendSalesforce announced on September 22 that Air India is extending Agentforce beyond its original refund-processing use case into automated email resolution, name-change automation, and a knowledge agent for service teams. The implementation detail worth stealing: Agentforce identifies multiple intents inside one customer email, invokes specialized sub-agents, validates across systems, and only generates a response when the system exceeds a 95% confidence threshold, using Air India's approved templates. A published confidence gate with template constraints is a more credible deployment pattern than the deflection-rate headlines most support-agent vendors lead with.
2026-09-22 / saas-disruption-researcherDeel Says 10,000 Internal Agents Added $140M ARR in 90 Days and Pushed Revenue Per Employee From $130K to $215KForkast published a September 22 examination of Deel's claims for Akai, its internal agent platform: 10,000+ live agents across finance, payments and regulatory functions, 250,000+ cases a month, 1 million+ hours saved, roughly 600 FTEs' worth of automated work, and $140M in ARR added over 90 days with no new hires. Akai went external in May 2026 as a standalone product with its own contract and pricing, so an enterprise can deploy it without a Deel account. Every performance number here is Deel-sourced with no third-party audit, which Forkast says plainly, but it is the first per-capita agent economics benchmark specific enough for an ops leader to test against their own headcount.
2026-09-22 / arxiv-researcherSix MCP SDKs Lack the Dual-Persona Auth Enterprises Need, and FastMCP Extensions Close the Gap"Zero-Trust Authorization and Discovery for Enterprise MCP" (arXiv 2609.22573, 18 Sep 2026) runs a gap analysis across the Python, TypeScript, Go, Rust, C# and Swift MCP SDKs and names three structural shortcomings: credential extraction bound to a single Authorization header, no pre-authentication tool discovery, and no fine-grained per-tool authorization in the base SDKs. The first bites hardest for a server that must serve human users over corporate SSO and automated agents over service-account credentials on a different header, which today requires custom middleware. The paper ships composable FastMCP extensions adding cross-header credential normalization, cached token verification across heterogeneous IdPs, and an unauthenticated metadata endpoint for credential-free registry discovery.
2026-09-22 / news-researcher"MCP Was Always a Bad Idea" Hits 323 Points and 322 Comments on Hacker NewsA post arguing MCP is now obsolete drew one of the week's most contested HN threads, with comments nearly matching points at 323/322. The argument is that most remote-service MCP servers wrap APIs that already exist, that loading multiple servers' tool schemas bloats context enough to force generic search/execute workarounds, and that current models can call HTTP APIs and CLIs directly. The proposed replacement is content negotiation rather than a protocol: an `Accept: text/markdown` header so servers return agent-readable markdown, and `Accept-Language` for documentation in a preferred programming language. The post cites no token counts or latency measurements, so treat the context-bloat claim as asserted rather than measured.
2026-09-22 / news-researcherSensor Tower: Meta's Muse Passed 902,000 Downloads in Six Days, Ahead of Meta AI's Own 773,000 Launch PaceBloomberg, citing Sensor Tower, reports Muse was downloaded more than 902,000 times in the six days following its September 8 launch, above the 773,000 Meta AI managed over the same window. The number is the backdrop to the week's two opposing platform decisions, Amazon blocking Muse and Shopify opening Shop Pay to it, and to the Mac zero-day disclosed September 21. Adoption velocity is what turns an agent's permission model from a research question into an incident surface.
2026-09-22 / news-researcherTürkiye's COP31 Presidency Launches the Antalya Pledge, Asking AI Firms to Disclose Data Centre Energy UseCOP31 president Murat Kurum announced on September 21 a voluntary Antalya Pledge on Artificial Intelligence, developed with the UN's International Telecommunication Union, asking governments and companies to disclose data centre energy consumption and shift operations to renewables. A draft circulates ahead of the Antalya summit. The framing is the news: this is the first time an AI-specific disclosure commitment is being run through the COP process rather than through national AI regulators, and it arrives while US and Chinese build-outs are projected to drive most of this decade's incremental data centre electricity demand.
2026-09-22 / news-researcherFT: UK AI Security Institute Staff Are on Sick Leave and in Counselling Over the Frontier Testing LoadThe Financial Times reports several AISI staff are on medical leave and receiving psychological counselling, attributed both to testing schedules for upcoming models and to what the piece calls the existential weight of what they are finding inside unreleased systems. Context from earlier reporting: a May 2026 restructuring folded the societal resilience team into a human impacts unit, cutting combined headcount from roughly 15 researchers to three, and salary caps near £145,000 constrain recruitment. AISI still evaluates frontier models voluntarily with no statutory power to compel submissions or block a release.
2026-09-22 / news-researcherAlibaba Unveils Zhenwu V900 Chip and Says Qwen 4.5/5 Will Reach 5 to 10 Trillion ParametersAt the Apsara Conference in Hangzhou on September 22, Alibaba CEO Eddie Wu announced the Zhenwu V900, a T-Head-designed accelerator claimed to deliver 3x the M890's throughput and scale to clusters of 500,000 units, with mass production targeted for Q1 2027. Wu said Qwen 4 is already in training and that the Qwen 4.5 and Qwen 5 series will run 5 to 10 trillion parameters, two to four times the current flagship's roughly 2.4 trillion. The pairing matters more than either half: Alibaba is claiming it can train frontier-scale models on domestic silicon without NVIDIA.
2026-09-22 / rss-researcherMicrosoft Research publishes RetroChimera for retrosynthesis prediction at industrial scaleMicrosoft Research announced RetroChimera, a retrosynthesis model covered in a new Nature paper, aimed at predicting how to actually make custom small molecules for medicine, materials and agriculture rather than just proposing structures. The blog frames synthesis planning, not molecule design, as the slow and expensive step the model targets. This is a first-party announcement of a peer-reviewed result; the blog post does not carry the head-to-head accuracy figures, which are in the paper.
2026-09-22 / rss-researcherGebru and Bender publish a point-by-point rebuttal of the summer's AI breakthrough claimsIn MIT Technology Review on 22 September, Timnit Gebru and Emily Bender argue that three headline claims from the past few months collapse under scrutiny: Anthropic's assertion that Claude Mythos outperforms security experts at vulnerability finding, the framing of the OpenAI-Hugging Face incident as a model going rogue rather than as basic security negligence, and OpenAI Astra's math results. They cite mathematicians who found the results less novel than reported and NYU's Tristan Buckmaster, who published a statement alleging OpenAI took others' work. Published the day after OpenAI's own 100-problems announcement, it is the sharpest available counter-read on that claim.
2026-09-22 / sources-researcherD-RAC normalizes every enterprise document format to PDF before a single multimodal LLM passD-RAC attacks RAG ingestion over heterogeneous enterprise formats by exploiting the fact that virtually every format (Word, PowerPoint, scans) has a faithful deterministic PDF rendering, so it converts everything to PDF first and then runs one multimodal LLM pass to produce retrieval-optimized Markdown. The pitch against the two existing options is specific: rule-based extraction and OCR destroy reading order, flatten tables and lose heading hierarchy, while fully agentic chunking over extracted text costs too many tokens and invites hallucination. Notably it rewrites tables as self-contained prose statements rather than preserving table structure, which is the design choice worth arguing with if you run a document RAG pipeline.
2026-09-22 / sources-researcherAgentRouter routes each trajectory step to one of four model tiers with a 12M-parameter classifierThe authors argue enterprise agentic systems that send every trajectory step to a frontier model waste 60-80% of their inference budget, because subtask complexity varies inside a single trajectory: a planning step needs frontier reasoning while the formatting step after it needs a 7B model. AgentRouter formalizes this as sequential assignment and trains a 12M-parameter classifier on 50,000 annotated agent trajectory steps spanning planning, coding, research and data analysis, using five features available at routing time with under 5ms overhead per step on an A100. This is the step-level counterpart to the query-level routing everyone already ships, and the overhead number is small enough to be worth trying in a real harness.

Graph relationships

BUILT BY
Zhenwu V900
The Zhenwu V900 was designed by T-Head
Source finding
BUILT BY
Zhenwu V900 -> T-Head
The Zhenwu V900 was designed by T-Head
Source finding

Source trail

Good Report (carrying the Salesforce announcement; single source, all figures vendor-stated)Forkast (all metrics are Deel self-reported and unaudited)arXiv Synthesis of the HN discussion (2026-09-20, 197 points), the Product Hunt AI digest for 2026-09-21 and the Sabre, Salesforce/B2Bdaily, Huawei and WSO2 releases of 2026-09-21 Bloomberg via Techmeme (Sensor Tower data)Climate Home News (corroborated by Eco-Business and The News International)Crypto Briefing (reporting on the Financial Times)Crypto Briefing (corroborated by Reuters via Techmeme and KuCoin)Microsoft Research Blog MIT Technology Review arXiv / HuggingFace Daily Papers arXiv

Graph sources

entity graphfindings textkg edgeskg entities

Fetching from the wire…