- AI KATANA
- Posts
- GPT-5 Tease Signals Leap Forward
In partnership with
51% of employees feel confused at work
Less than half say they can easily find the...
goals
strategies
directives
...their leaders have shared. Worse, only 9% of staff feel tightly aligned with org-wide goals.
We surveyed 1,200+ professionals to see what's going wrong — and how to get it right.
Hello!
Another 24-hour news sprint and the AI world hasn’t paused for breath. OpenAI’s next-gen model is suddenly real enough for Sam Altman to hint at on X; telecom giant Bharti Airtel is standing up a pan-Asian cloud packed with AI services; Broadcom just delivered a monster switch that can knit GPU clusters 100 km apart; Washington floated a bipartisan “TEST AI Act” to standardise safety testbeds; DXC and 7AI unveiled an agent-driven SOC built around autonomous incident responders; and Palantir’s soaring commercial demand forced it to raise revenue guidance—again. Toolmakers kept pace: Experian rolled out a compliance-minded model-risk assistant and D-Wave open-sourced a quantum-AI toolkit that plugs straight into PyTorch. On the money front, London-based CuspAI is courting a nine-figure round for materials discovery while Seoul’s PhnyX Lab quietly nabbed seed capital to bring GenAI into pharma search. And because even insiders need a chuckle, Google’s new Pixel ad skewers Apple for AI features that are still “coming soon.”
Sliced just for you:
🔮 GPT-5 Tease Signals Leap Forward
☁️ Airtel Rolls Out AI Cloud
🚀 Broadcom’s Jericho4 Super-Switch
🏛️ TEST AI Act Hits Senate
🛡️ DXC & 7AI Agentic SOC
📊 Palantir Ups Forecast on AI Boom
In a single cryptic post on X, Sam Altman invited GPT-4 to define an “impossible shape,” then answered his own riddle by revealing he’d asked GPT-5 first—implying the new model is already training and producing outputs that unsettle its creator. Observers parsed the exchange and discovered tell-tale log lines showing throughput well above the 900-token context limit that binds GPT-4o mini, hinting at a fundamentally larger architecture with long-term memory and upgraded reasoning heuristics. Leaked dev-day decks circulating among beta partners describe a modular system able to hot-swap specialised “skill adapters” without full retraining, plus tooling for verifiable provenance that may satisfy emerging U.S. and EU disclosure rules. Altman’s teaser suggests a public preview at OpenAI’s November conference—earlier than many analysts expected.
Bharti Airtel’s digital arm Xtelify unwrapped a regional cloud and software stack built for operators and large enterprises across India, Singapore, the Philippines and Africa. The Infrastructure-as-a-Service layer pairs local data-sovereignty zones with GPUs optimised for GenAI training, while a companion telco suite offers AI-driven workforce routing, churn prediction and real-time ARPU boosters. Launch partners Singtel, Globe Telecom and Airtel Africa will co-market the platform, giving Airtel an end-to-end play against AWS and Azure in markets where sub-second latency is critical and undersea cables scarce. Analysts note the addressable Indian cloud market is projected to triple to $24 billion by 2028, and Airtel’s 360 million-strong mobile base provides immediate distribution for AI APIs.
Broadcom debuted Jericho4, a 3-nm router-on-chip that can mesh thousands of AI servers across 60-mile “metro clusters” without bottlenecks. By embedding high-bandwidth memory directly on the switch and adding in-flight encryption, Jericho4 moves 16 Tbit/s per device while reducing packet loss under heavy GPU synchronisation loads—crucial for large-scale model training. A single petascale system could chain 4,500 chips, enough to keep pace with Nvidia’s next-wave accelerators. Cloud giants eye the design to sidestep scarce long-haul fiber and future-proof data-centre security. Industry watchers say the part will ship in Q1 2026 and could blunt start-ups touting optical interconnects.
New bipartisan bill S.1633 (Testing and Evaluation Systems for Trusted AI Act) tasks NIST with building open testbeds that measure robustness, bias and safety for advanced systems before federal adoption. The Act would inject $800 million over five years, mandate red-team access for researchers, and require periodic public scoring akin to the National Vulnerability Database. Sponsors argue voluntary benchmarks leave a gap between lab claims and real-world performance; critics warn fast-moving private models could outstrip static government suites. Hearings are slated for September with strong defence-committee interest following DARPA’s data-poisoning alerts.
DXC Technology paired with start-up 7AI to launch an “Agentic SOC” that leans on swarms of autonomous detection agents coordinating via a large-language control plane. Early pilots saw incident triage times drop from 17 minutes to 90 seconds and saved clients 224,000 analyst hours. The service pipelines 7AI’s reasoning engines into DXC’s Platform X, letting customers plug custom playbooks while keeping data on-prem. Insurance and healthcare clients are first in line as they juggle tight breach-notification windows and rising ransomware. DXC plans to franchise regional nodes under a revenue-share model by year-end.
Data-analytics heavyweight Palantir lifted its 2025 revenue outlook to as much as $4.15 billion—its second upward revision this year—after U.S. commercial sales of its AI Platform jumped 62 % YoY. Government pipelines remain strong, but management highlighted a shift: more than 57 % of new bookings now come from private-sector deployments in logistics, insurance and energy. Margin guidance also widened thanks to reusable AI inference templates that cut delivery costs. Investors pushed shares up 5 % in late trading, convinced the company’s pivot away from customised dashboards toward packaged AI “modules” is sticking.
🛠️ AI tools updates
Credit-bureau Experian added an AI assistant to its Ascend platform that automates model-risk documentation, validation and ongoing surveillance. Banks piloting the tool report internal approval cycles shrinking by 70 %, aligning with Fed SR 11-7 and UK PRA SS1/23 guidance. Built-in generative explainers create audit-ready narratives, and drift alerts flag credit models that wander off spec. The launch targets compliance teams weary of spreadsheet-driven governance as regulators dial up scrutiny of black-box lending algorithms.
Quantum-computing pioneer D-Wave released an open-source toolkit that wires its annealing processors directly into PyTorch workflows. Developers can now train restricted Boltzmann machines partly on-quantum hardware, accelerating convergence on generative tasks like molecular design. The bundle ships with demos and a LaunchPad programme offering subsidised QPU credits, signalling D-Wave’s push to blend classical and quantum resources for real-world AI workloads.
💵 Venture Capital updates
Cambridge-born CuspAI is negotiating a raise “north of $100 million” to scale its generative-design platform that searches chemical space for carbon-negative materials and next-gen batteries. The round would follow last year’s $36 million seed and could value the firm above $700 million, giving Europe a rare deep-tech play ready to rival U.S. incumbents. Proceeds are earmarked for supercomputer time and wet-lab validation partnerships.
Seoul-based PhnyX Lab closed a $4 million seed led by SK Networks with angel cheques from Cohere’s Aidan Gomez and NEAR’s Illia Polosukhin. Funds accelerate its Cheiron suite, which parses medical literature and drafts clinical protocols with GenAI co-pilots. Over 60 pharma clients already use Cheiron to shave weeks off trial-design cycles; fresh cash expands into automated regulatory writing and multilingual evidence aggregation.
🫡 Meme of the day
⭐️ Generative AI image of the day
Before you go, check out Google Roasts Apple’s “Coming Soon” AI.
