> SPOTLIGHT
WHAT MATTERS TODAY
OpenAI says GPT-Live uses a full-duplex architecture. It can keep receiving input while generating output, deciding whether to speak, keep listening, pause, interrupt or call a tool. When a request needs search, deeper reasoning or longer work, GPT-Live can delegate that work to a frontier model such as GPT-5.5 in the background while keeping the conversation open.
OpenAI describes two architectural changes. Older voice systems chained speech-to-text, a language model and text-to-speech, while turn-based voice models still waited for the user to finish speaking. GPT-Live processes input and output continuously, then brings background search or reasoning back into the same conversation.
OpenAI tested the system in two ways. Direct user comparisons measured turn-taking, interruptions, conversational flow and naturalness. Specialist tests measured expert scientific reasoning through GPQA, agentic web search through BrowseComp, and multi-turn support work through τ³-Voice Telecom.
In its Q2 shareholder letter, Palantir argues that enterprises should retain control of their data, prompts and business intelligence instead of handing the operating layer to language-model companies. The letter also reports roughly $1.9 billion in quarterly revenue, up 93% year over year, and roughly $1.1 billion in profit.
Palantir describes sovereignty through five moves: own, route, govern, deploy and learn. That means keeping enterprise context portable across models and environments, routing models according to value, risk and continuity, applying permissions before an agent reads, reasons or acts, and recording prompts, tool calls, evaluations, approvals and results in an audit trail.
Palantir also asks what happens when a model provider changes access, retention, pricing or model behavior. Its answer is to keep data, code, vulnerabilities, policies and tradecraft as enterprise context, measure task quality before and after model swaps, and maintain failover across model families or hosting surfaces.
Kimi K3 raises a question underneath the benchmark race: how should a model remember and serve a context window that reaches 1 million tokens? SemiAnalysis examines the architecture behind that question. Kimi K3 has 2.8 trillion parameters, around 104 billion activated parameters, and uses two forms of attention. Kimi Delta Attention, or KDA, keeps a fixed-size memory state, while full-attention layers can still look back across the sequence when needed. The model also has 896 experts, with 16 activated for each token.
This design means the serving system has to manage two kinds of memory. One stores token-level information for full attention, while the other stores the compact state used by KDA. Reusing processed context must account for both. vLLM also describes separate work for reading long context, generating responses, distributing work across GPUs, predicting tokens and checking them, and calling tools.
SemiAnalysis and vLLM report that KDA reduces memory growth associated with longer context, while making the serving stack more complex. In vLLM's benchmark, Kimi K3 reached 118 tokens per second without speculative decoding and 370 tokens per second with DSpark on 16 NVIDIA GB300 NVL72 GPUs. Those results apply to that reported configuration.
> PRESENTED BY OMNIPHI
Agents Came to Software. Now Trading.
You have seen what coding agents changed. You describe a goal, they build it, and you review the work. That shift is reaching trading.
OmniPhi brings charts, research, backtesting, code, news, and broker connections into one workspace. At the center is Phi, an agentic harness that builds what you describe in plain English.
Ask for a scanner, a news watcher, or a full strategy. Phi researches it, builds it, tests it against historical data, and runs it live on your own broker account, within the guardrails and intent you set up front. Run one agent or several, each with its own job.
This is the Agentic Integrated Trading Environment. Live on OANDA, Alpaca, Kraken, and Binance, with more coming soon. The beta is open, and every application is read by hand.
Disclaimer: OmniPhi is a software tool, not investment advice. Trading carries substantial risk of loss. Phi executes within the guardrails and intent you set up front.
> SIGNAL HEADLINES
CAPTURE THE SHIFT
Vibe coding goes private: AWS and Superblocks announced a multiyear agreement to bring vibe coding into customers' private clouds, with data, applications and inference inside the enterprise boundary. Superblocks describes a deployment model that runs inside a customer's AWS VPC, sends inference through the customer's Amazon Bedrock account, and uses models selected by the organization.
Horizon3 raises on AI security: TechCrunch reports that Horizon3 raised a $250 million Series E at a $2 billion valuation as the article examines how AI is expanding the attack surface. The round was reported as led by NightDragon and NEA, with Horizon3 described as using AI for security testing.
AMD warrants reprice GPU economics: SemiAnalysis models how performance-based warrants between AMD, Meta and OpenAI could reduce modeled capital cost per GPU-hour under different share-price and deployment assumptions. The analysis reports modeled costs, not cash savings confirmed by the companies.
> ONE PRACTICAL USE OF AI TODAY
Find the Failure Point
An agent is a system made up of a model, a harness that coordinates prompts, context and tools, plus users, tools, memory and the environment. “The agent did not complete the task” is therefore not enough to conclude that the model's reasoning failed.
This classification separates 41 failure modes by the point of contact between two components and identifies which side needs to be fixed.
The five main points are:
Model and harness: the model misunderstands the instruction, or the coordination layer supplies incomplete context, excessive context or the wrong loop.
Harness and tool: the schema, parameters, tool result or retry logic causes the agent to call the tool incorrectly or miss the result.
Tool and environment: permissions, network, files, API state or outside changes create a different real-world state than expected.
User and memory: memory is stored incorrectly, retrieved in the wrong context or carried over from an older task.
Environment and grader: the agent behaves correctly but the environment or benchmark records the result incorrectly.
The classification connects each failure to a repair path. Model-side errors may call for post-training, harness errors for scaffolding or tool-integration work, environment errors for runtime changes, and grader errors for a redesigned benchmark.
> WORTH READING
ANALYSIS & THESIS
MIT Sloan proposes six questions for an AI strategy: what shared ambition the organization is pursuing, how governance will work, how value can scale, whether the technical foundation is ready, whether the culture supports a new way of working, and which skills the team needs.
The six areas connect parts of an AI program that are often treated separately. Shared ambition sets the common goal. Governance defines decision rights and accountability. Scaling value asks how successful use cases can expand. Technical foundation covers data and infrastructure. Culture covers how the organization absorbs change. Skills define the capabilities needed to build, operate and manage AI systems.
TechRadar describes agents as a layer between software and the user. They can interpret objectives, plan, choose tools, access data and take actions. The article carries familiar security principles such as least privilege, authentication and separation of duties into systems that can act on a user's behalf.
It argues that each agent needs its own identity, owner, purpose, limited permissions and lifecycle with review points and an expiry date or recertification. The audit trail should preserve both the person who initiated or authorized an action and the agent that executed it. Organizations also need an inventory of agents across SaaS, internal automation, development environments and third-party integrations.
