Index Company
43 CÈ pieces since 11 May 2026: 30 intel items, 13 essays. Newest first. All companies →
-
Tests cannot cover what an agent will try: Google open-sources a Z3 prover for CEL policies
Google published a formal verification framework for Common Expression Language in the cel-java repository. It compiles policies into SMT and uses the Z3 theorem prover to check every input, catching gaps such as a port slipping past a malformed range rule.
-
The model was never the moat: DeepSeek open-sources its agent harness under MIT
DeepSeek published Harness v0.1 under the MIT licence, built on the Cordis meta-framework where models, tools, sandboxes, sessions, and orchestration are all plugins. It ships OS-level sandboxing and supports Anthropic, OpenAI, Bedrock, Azure, and Gemini.
-
Google’s SEA accelerator gives 25 founders $350,000 in credits and takes three months.
Memento mori, applied to an acceptance email: the cohort takes no equity, and the one thing it does take is the only asset you cannot raise more of.
-
Cheap intelligence with a return date: Gemini 3.7 Flash costs $0.75 until December 31
Google shipped Gemini 3.7 Flash at $0.75 per million input tokens and $3.75 output, both introductory rates that double on January 1, 2027. FrontierCode 1.1 scores rose to 43.6% from 34.4%.
-
Alphabet's first cash burn since 2004: $44.9B of quarterly capex turns free cash flow negative
Alphabet posted negative free cash flow of $5.85 billion in the second quarter, its first as a public company, after capital spending doubled to $44.9 billion and full-year guidance rose to a range of $195 to $205 billion.
-
Sovereign AI in Southeast Asia (running on Hangzhou’s open weights)
Google says close to seven in ten Southeast Asian prompts now arrive in a local language, and the region has its own models to match; read the model cards and the sovereign part turns out to be the last layer.
-
When intelligence keeps repricing downward, who captures the saving? Gemini 3.6 Flash cuts output to $7.50 per million tokens
Google released Gemini 3.6 Flash on July 21 at $1.50 per million input and $7.50 output tokens, below its own 3.5 Flash, while burning 17% fewer output tokens and extending its knowledge cutoff to March 2026.
-
Malaysia’s draft AI law treats your training data and your model’s output as property
The National AI Office’s consultation closes 31 July, and the clause the Prime Minister calls the heart of the bill would let Kuala Lumpur, rather than your terms of service, decide who owns what a model trains on and what it generates.
-
Four frontier models shipped in ten weeks. From high enough, not one of them happened.
On the view from above, and the summer four frontier models shipped while a founder built nothing.
-
Indonesia sentenced Gojek’s co-founder to ten years over a laptop procurement
On 30 June a Jakarta court sent Nadiem Makarim to prison for ten years without finding he took a cent, and the part every Southeast Asian founder should study is how a lawful call became a crime.
-
The biggest customer the AI labs have is quietly building their replacement
Microsoft began routing tens of thousands of Excel and Outlook prompts a week to its in-house MAI models this month. The question that leaves for anyone building on AI is not which lab to trust.
-
Salesforce ships Agentforce Commerce with agents that buy and sell inside ChatGPT
On July 6 Salesforce made its Shopper, Buyer, and Merchant agents generally available, with native integration into ChatGPT and Google's AI Mode and Gemini to follow, moving retail transactions from search results into agent-mediated chat.
-
What an AI agent can actually run in an SME back office right now
Two frontier models shipped in five weeks and the real product is admin, not code. The workflows an agent finishes alone are the ones that never leave the ledger; the ones that break are the ones that touch a person.
-
Anthropic, OpenAI, and Google converge on the agent harness as product
The runtime that wraps a model with memory, tools, and control loops is now the sold unit. Anthropic prices Managed Agents at eight cents per session hour; OpenAI ships its harness as an open-source Agents SDK update.
-
Southeast Asia now holds 2,000 data centres. It captures almost none of the margin.
The land, power and water are local; the models, the margin and the IP are almost entirely foreign, which makes the operator question simple: own a customer, not a GPU.
-
Google DeepMind ships DiffusionGemma, the first open-weight text-diffusion model
The 26B Apache-2.0 model denoises 256-token blocks in parallel rather than one token at a time, clearing 1,000 tokens per second on a single H100, roughly four times faster than autoregressive Gemma.
-
OpenCode reaches 160,000 stars as 7.5 million developers adopt model-agnostic coding
The open-source CLI agent connects more than seventy-five providers, from Claude and GPT to Gemini and local Ollama models, through a single interface, and now serves 7.5 million monthly developers.
-
The cheap AI on your desk is priced below cost
Four giants are losing money on every token to win the market, the bill for the buildout runs to US$5.5 trillion, and Singapore’s own forecasters have started naming the morning the subsidy stops.
-
Google donates A2A protocol to Linux Foundation for agent interoperability
Google donated the A2A protocol to the Linux Foundation on June 23, 2025, establishing it as a vendor-neutral standard for AI agent communication across framework boundaries, with founding members including AWS, Cisco, Microsoft, Salesforce, SAP, and ServiceNow.
-
Apache Iceberg 1.11.0 adds server-side scan planning and Google KMS encryption
Apache Iceberg 1.11.0 launched with support for Apache Spark 4.1 and Flink 2.1, REST catalog server-side scan planning, partition statistics API, built-in table encryption with Google KMS, and Google Storage Analytics integration.
-
Google unveils TurboQuant algorithm cutting KV cache memory overhead by six times
Presented at ICLR 2026, TurboQuant uses PolarQuant vector rotation and Quantized Johnson-Lindenstrauss compression to reduce the key-value cache bottleneck, enabling massive-context models to run far more efficiently.
-
Google donates A2A agent-interop protocol to Linux Foundation under neutral governance
Announced at Open Source Summit North America in June 2025, the A2A protocol became vendor-neutral in 2026 with founding members including AWS, Cisco, Microsoft, Salesforce, SAP, and ServiceNow. It enables agents to communicate across framework boundaries.
-
From a Singapore HQ on a Monday: what Shopee’s cuts say about knowledge work
Sea cut roughly 8 percent of Shopee’s developers and blamed AI by name. The number that actually moved was the cost of shipping the next thing, and headcount followed it down.
-
Google ships Gemma 4 12B under Apache 2.0, ending custom AI license era
Released June 3, Gemma 4 12B is Google's first multimodal model family under the permissive Apache 2.0 license, replacing the restrictive Gemma Terms of Use that enterprise legal teams flagged as commercially unworkable. The 12-billion-parameter model runs on any laptop with 16GB RAM.
-
Apache Iceberg one-point-eleven-zero ships server-side scan planning and table encryption
Iceberg 1.11.0 adds Spark 4.1 and Flink 2.1 support, REST catalog server-side scan planning, partition statistics API, built-in envelope encryption, and Google KMS integration.
-
Apple ships LanguageModel protocol letting iOS apps swap Foundation, Gemini, Claude without code rewrites
The new Swift Package Manager protocol announced at WWDC 2026 lets developers switch between Apple Foundation Models, Google Gemini, and Anthropic Claude with no session-code changes, while Xcode 27 ships agentic coding features and multi-turn streaming in the first beta.
-
Google donates A2A protocol to Linux Foundation, forming A2A Project with AWS, Cisco, Microsoft, Salesforce, SAP, ServiceNow
On June 23, 2025, Google officially donated the A2A protocol to the Linux Foundation at Open Source Summit North America in Denver, establishing it as a vendor-neutral standard for AI agent interoperability, with founding members including Amazon Web Services, Cisco, Microsoft, Salesforce, SAP, and ServiceNow.
-
Apache Iceberg 1.11.0 ships server-side scan planning, Google KMS encryption, Spark 4.1 and Flink 2.1 support
Apache Iceberg 1.11.0 adds support for Apache Spark 4.1 and Apache Flink 2.1 as default build targets, while the REST catalog learns to plan scans server-side, a new partition statistics scan API gives optimizers clean access to table shape, and built-in table encryption arrives with envelope encryption and Google KMS support.
-
Apache Iceberg 1.11.0 ships production-ready V3 spec with deletion vectors and table encryption
The May 19 release adds server-side scan planning in REST catalogs, native encryption with Google KMS support, and stabilizes V3 features including deletion vectors and nanosecond timestamps, while upgrading default support to Spark 4.1 and Flink 2.1.
-
MiniMax M3 ships one-million-token coding model but withholds training code
Chinese startup MiniMax launched M3 on June 1, 2026 with claims of topping GPT-5.5 and Gemini 3.1 Pro on benchmarks, yet has not released training code or inference operators, leaving it in open-weight territory rather than fully open source.
-
Apache Iceberg 1.11.0 ships server-side scan planning and Google KMS envelope encryption
Iceberg 1.11.0 launched with support for Apache Spark 4.1 and Flink 2.1 as default build targets, REST catalog server-side scan planning, a new partition statistics scan API, and built-in table encryption with envelope encryption and Google KMS support, accelerating Iceberg workloads on Google Storage Analytics library integration.
-
Open Design hits 57,400 GitHub stars as local-first alternative to Claude Design
Open Design reached v0.9.0 on June 2 with 310 contributors, offering model-agnostic design artifact generation, self-hosting, and data residency control. The tool auto-detects local coding agents, stores everything in SQLite, and supports Claude, OpenAI, Gemini, and Ollama endpoints with BYOK proxy and SSRF protection.
-
Google donates Agent-to-Agent protocol to Linux Foundation with AWS, Cisco, Microsoft, Salesforce backing
Google officially donated the A2A protocol to the Linux Foundation on June 23, 2025, establishing it as a vendor-neutral standard for AI agent interoperability. Founding members include Amazon Web Services, Cisco, Microsoft, Salesforce, SAP, and ServiceNow, enabling agents to communicate across framework and organizational boundaries without single-provider lock-in.
-
Apple opens Foundation Models framework to third-party models and free Private Cloud Compute for small developers
At WWDC Platforms State of the Union on June 9, Apple announced free Private Cloud Compute access for developers with under two million App Store downloads, image input support, server-side model integration for Claude and Gemini, and Dynamic Profiles for multi-agent workflows. The framework will go open source this summer.
-
What it means when an LLM becomes the gatekeeper of brand discovery in Southeast Asia
Across APAC, more than half of consumers now run purchase decisions through an AI assistant, and the assistant defaults to whoever is most legible to it. That is a strategic problem dressed as a marketing one.
-
Google ships Gemma 4 12B multimodal model under Apache 2.0, runs on 16GB laptops
Released June 3, Gemma 4 12B processes text, images, audio, and video with an encoder-free architecture using 11.95 billion parameters, available under Apache 2.0 for unrestricted commercial use and running entirely on standard enterprise laptops.
-
Apache Iceberg 1.11 shifts metadata orchestration server-side, unifies cloud catalogs
Iceberg 1.11.0 pushes scan planning into the REST catalog and adds native support for Spark 4.1 and Flink 2.1, while Google's Cloud Storage Analytics integration cuts Parquet metadata latency across multi-cloud deployments.
-
Open Design reaches 57.4K GitHub stars in eight weeks as self-hosted Claude alternative
Community project Open Design hit v0.9.0 on 2 June with 310 contributors, 1,837 commits, 6,500 forks. Offers local-first design-artifact generation, multi-provider API proxy (Anthropic, OpenAI, Azure, Gemini, Ollama), and Claude Design ZIP import. Runs via pnpm, deploys to Vercel, stores in SQLite.
-
Google releases Gemma 4 under Apache 2.0, first OSI-approved open license in Gemma family
Gemma 4, announced at I/O 2026, is the first Gemmaverse model under Apache 2.0, enabling commercial modification and redistribution without proprietary restrictions. Downloaded over 500 million times since the family launched in 2024, Gemma now powers sovereign AI in Ukraine, India's 22-language Project Navarasa.
-
Google ships Gemini 3.5 Flash at four-times frontier speed, half the API cost
Released May 19, Gemini 3.5 Flash beats last year's Pro tier on coding and agentic benchmarks while running 280 tokens/second—priced at $1.50 in, $9 out per million tokens, roughly 40% cheaper than Gemini 3.1 Pro.
-
Google releases Gemma 4 models under OSI-approved Apache 2.0, first true open-source Gemma
Gemma 4 launched in April 2026 as the first Gemmaverse release under Apache 2.0 license, enabling developers to modify, build, and distribute derivatives freely—already supporting sovereign infrastructure from Ukraine state licensing to India's 22-language Project Navarasa.
-
MTI raised the 2026 forecast. Manufacturing had its worst quarter since 2023.
Singapore posted Q1 GDP up 4.6 percent year on year. Manufacturing fell 4.9 percent quarter on quarter, the steepest single-quarter drop since 2023. The AI hardware tailwind has rolled off.
-
What a ten-person Singapore startup will pay Microsoft to govern its AI agents.
Microsoft moved Agent 365 to general availability on May 1, priced at fifteen US dollars per user per month. The governance layer is the small number. The compute it brackets is the large one.