OpenAI collapsed its fragmented model lineup into GPT-5.4, the first model to combine reasoning, coding, and computer use while simultaneously shipping Codex-Spark at 1,000+ tokens per second on Cerebras hardware, a quiet signal that inference infrastructure is diversifying beyond NVIDIA. Amazon lost 6.3 million orders in a week to AI-linked outages, then mandated a 90-day safety freeze across 335 critical systems, the starkest enterprise signal yet that governance hasn’t kept pace with adoption. Cursor, Anthropic, and Ramp each shipped a different piece of the agentic infrastructure layer in the same cycle. Meta delayed Avocado to at least May after internal tests placed it between Gemini 2.5 and 3.0. And the MCP debate came to a head when Perplexity’s CTO publicly abandoned it for plain APIs and CLIs. Outside the technical stack: the Pope used his final public address to remind every leader that human dignity should constrain technological ambition, and Meta’s Ray-Ban smart glasses sparked a privacy controversy after workers revealed they can see everything the wearer does in real time.
GPT-5.4 ships as the first model to combine coding, reasoning, and computer use. OpenAI collapses what used to be three specialized variants into one.
Amazon’s AI coding tool contributed to 6.3M lost orders in a single week. The company is now enforcing a two-person review and a 90-day freeze on 335 critical systems.
In the same cycle, Cursor launched always-on automation agents, Anthropic shipped PR review at $15–25 per review, and Ramp introduced payment cards built natively for AI agents. The agentic infrastructure layer is assembling faster than any single vendor can own it.
OpenAI’s Codex-Spark hits 1,000+ tokens/sec on Cerebras hardware while vLLM ships 3-path AMD routing at 4.4x legacy throughput, inference is splitting into workload-specific tiers, and NVIDIA’s monopoly on AI serving is softening.
Meta spent $115–135B on AI infrastructure, and Avocado still landed between Gemini 2.5 and 3.0; compute spend and frontier model output are decoupling.
GPT-5.4 merges the frontier coding capabilities of GPT-5.3-Codex into a general-purpose reasoning model for the first time. There was no GPT-5.3 Thinking, making this a significant capability jump.
Key numbers: 83% on GDPval (up from 70.9% for GPT-5.2), 83.3% on ARC-AGI-2 for the Pro variant (vs. 54.2%), 33% fewer false individual claims and 18% fewer error-containing responses versus GPT-5.2. In the API, a new Tool Search system loads tool definitions on demand rather than upfront, cutting token consumption by 47% in tested scenarios, and the model supports up to 1 million tokens of context in Codex. The system card designates GPT-5.4 Thinking as the first general-purpose model with active cybersecurity-capability mitigations under OpenAI’s Preparedness Framework, a shift from flagging the risk to building guardrails for it. Bringing computer use, frontier coding, and reasoning into a single model removes the need to route between specialized variants, which will pressure competing labs to consolidate their own fragmented model lines.
Notably, roughly 90% of Codex itself is now written by Codex, per the team. The Cerebras partnership is OpenAI’s first production signal that inference hardware is diversifying beyond NVIDIA, and that latency-optimized chips will serve a distinct tier of AI workloads.
Two incidents in early March 2026 triggered Amazon’s response: on March 2, Amazon’s AI coding assistant Q contributed to an outage that produced 1.6 million website errors and 120,000 lost orders; on March 5, a separate event caused a 99% drop in North American marketplace orders (6.3 million lost orders) after a production change was deployed without the required Modeled Change Management process.
The context: a November 2025 internal memo signed by two SVPs had mandated Kiro as Amazon’s standard AI coding tool with an 80% weekly usage target set as a corporate OKR, and internal dashboards tracked whether engineers hit minimum daily AI usage targets, incentivizing volume of AI-assisted code at exactly the moment governance hadn’t caught up. SVP Dave Treadwell acknowledged in a memo that “best practices and safeguards” around generative AI usage are not yet fully established, describing a pattern with “high blast radius” tied to “Gen-AI assisted changes.”
The 90-day reset mandates a two-person review for all code changes to the 335 highest-priority systems, formal documentation before deployment, and automated pre-deployment validation. Amazon disputes that AI wrote the faulty code directly, framing these as “user error”, but the combination of mandatory AI adoption targets and immature deployment guardrails tells a more structural story.
Meta’s next-generation model, codenamed Avocado, has slipped from March to at least May, its third delay, having originally targeted late 2025 after internal benchmarks placed it between Google’s Gemini 2.5 and Gemini 3.0, the latter of which launched in November 2025. Releasing a model that trails a four-month-old competitor’s release would be a difficult position for a company that publicly committed $115–135 billion in AI capital spending for 2026 alone. The performance gap is specific: shortfalls in reasoning, coding, writing, and agentic behavior, exactly the capabilities Meta’s restructuring was supposed to fix.
The organizational context makes the delay more significant than a typical model slip. In June 2025, Meta invested $14.3 billion in Scale AI and installed founder Alexandr Wang as Chief AI Officer, explicitly promising investors the highest talent density in the industry. Wang built TBD Lab (a ~100-person internal division developing Avocado (text/reasoning) and Mango (image/video)) while hundreds of researchers from Meta’s existing FAIR unit were laid off and Yann LeCun departed to found AMI Labs. Ruoming Pang, recruited with a $200M package, left for OpenAI after just seven months. Internal tensions between Wang, CPO Chris Cox, and CTO Andrew Bosworth, particularly over whether Avocado should prioritize advertising performance or frontier capability, have compounded the execution challenge.
The most striking detail in the NYT report: Meta’s AI leadership discussed temporarily licensing Google’s Gemini to power Meta AI products across WhatsApp, Instagram, and Facebook while Avocado catches up. No decision has been confirmed, but the possibility alone, Meta licensing technology from its primary rival in advertising, smart glasses, and AI, signals how much pressure the company is under. Fortune called it “almost impossible to imagine.” Whether Avocado will launch as open or closed source also remains unresolved, a question that carries significant strategic weight given Meta’s historical Llama identity. What’s next in the pipeline: a model codenamed Watermelon. The gap between frontier leaders and fast followers is proving more durable than Meta’s infrastructure spending assumed.
Claude Code’s run-rate revenue has exceeded $2.5 billion since launch, and the root cause of the new product is simple: Claude Code dramatically increased code output, but PR review velocity didn’t scale with it, creating a bottleneck at the merge step. Code Review addresses this with multiple Claude agents analyzing a PR in parallel from different perspectives, with a final agent aggregating and ranking findings by severity (red/yellow/purple), focused specifically on logical errors rather than style. The tool integrates directly with GitHub, leaves comments inline, and is available in research preview to Teams and Enterprise customers at an estimated $15–25 per review. The same company generating the code is now selling the review layer, and as agents produce more code, the market for AI-powered quality gates will expand proportionally.
Karpathy’s Autoresearch is a Python script that runs AI agents overnight to autonomously generate hypotheses, run training experiments, evaluate results, and iterate. The experiment run shown produced 15 kept improvements to a language model training script, reducing validation BPB from ~1.000 to ~0.975 without human intervention. Each experiment runs on a feature branch, the agent merges improvements when they lower validation loss, and the entire process is visualized as a running best-loss curve with labeled interventions. The project is explicitly framed as “part code, part sci-fi, and a pinch of psychosis” — a proof-of-concept rather than production tooling — but it demonstrates a self-improving research loop on commodity hardware that foreshadows what automated ML research infrastructure will look like at scale.
Gemini 3.1 Flash-Lite is positioned as the throughput-first option in Google’s model lineup, outputting 363 tokens per second compared to 249 for Gemini 2.5 Flash, 71 for GPT-5 mini, and 108 for Claude 4.5 Haiku. On benchmark comparisons, it scores 86.9% on GPQA Diamond, 76.8% on MMMU-Pro, 84.8% on Video-MMMU, and 88.9% on MMLU, competitive with or exceeding comparably priced models in most categories, with the exception of FACTS Benchmark (40.6% vs Gemini 2.5 Flash’s 50.4%) and SimpleQA (43.3%).
Pricing at $0.25 input and $1.50 output per million tokens undercuts Claude 4.5 Haiku ($1.00/$5.00) significantly, maintaining price parity with GPT-5 mini on input while offering substantially higher throughput. The MRCR v2 long-context score of 12.3% at 1M context marks the only area where it materially trails Gemini 2.5 Flash (21.0%), signaling that Flash-Lite trades long-context performance for speed and cost efficiency.
The benchmark comparison table from the Gemini 3.1 Flash-Lite model card showing scores, pricing, and output speed across Gemini 3.1 Flash-Lite, Gemini 2.5 Flash, Gemini 2.5 Flash-Lite, GPT-5 mini, Claude 4.5 Haiku, and Grok 4.1 Fast, best illustrates the full cost-capability-speed tradeoff.
Phi-4-reasoning-vision-15B uses a mid-fusion architecture with a SigLIP-2 Naflex dynamic-resolution encoder over a Phi-4-Reasoning backbone, achieving 88.2% on ScreenSpot-v2 (GUI grounding), 75.2% on MathVista-Mini, and 83.3% on ChartQA — while requiring roughly one-tenth the compute of comparably-capable models like Qwen3-VL-32B at inference time. The model is trained on a 20/80 mix of reasoning and non-reasoning data, defaulting to direct responses for perception tasks and invoking chain-of-thought only for math and science, reducing unnecessary verbosity and latency. Notably, the full multimodal training used just 200 billion tokens, versus 1+ trillion for Qwen 2.5 VL, Kimi-VL, and Gemma3. The release includes weights, fine-tuning code, and full evaluation logs, and signals that the efficiency gains from the Phi data-quality philosophy are extending cleanly from language-only to multimodal reasoning.
Ramp Agent Cards provide AI agents with the ability to make purchases governed by merchant controls, real spend limits, and full transaction visibility, addressing what Ramp describes as the absence of any safe financial mechanism for agentic systems. The product is the first fintech offering purpose-built for agent-initiated spend rather than retrofitting human card infrastructure, and it signals that financial infrastructure companies are treating agent autonomy as a distinct product category requiring native tooling. As AI agents gain the ability to browse, purchase, and transact independently, payment rails and spend governance will become foundational infrastructure requirements alongside compute and memory.
Cortical Labs’ CL1 system embeds 200,000 living neurons grown from adult human skin and blood samples onto a silicon chip, connected to a real-time feedback loop, training time to play a 3D shooter dropped from 18 months on earlier hardware to one week. A 30-unit server rack draws only 850–1,000 watts total, versus megawatts for a large GPU training cluster; each unit costs $35,000, and the company is now selling compute access via “Wetware as a Service” through Cortical Cloud, allowing developers to deploy code to living neurons remotely with a software subscription model. The CIA’s In-Q-Tel is among the backers, and 115 units began shipping in 2025. The biological compute case is not raw speed but energy efficiency and adaptive learning under uncertainty; the gap between silicon and biological compute on those specific axes may widen as models scale.
Cursor Automations allows teams to configure cloud-hosted agents that spin up sandboxes, execute instructions using configured MCPs and models, verify their own output, and learn from past runs via a persistent memory tool, removing the human trigger requirement entirely.
Pre-built templates cover security review on every push to main, agentic PR risk classification with auto-approval for low-risk changes, incident response triggered by PagerDuty with Datadog log analysis, and morning test-coverage checks on recently merged code. Cursor’s BugBot, which now runs thousands of times daily and has caught millions of bugs, is described as the original automation pattern that Automations generalizes. This shifts Cursor’s product position from IDE assistant to continuous engineering infrastructure: a “factory that creates your software”, competing directly with purpose-built CI/CD and DevOps automation platforms.
Sakana AI trained a 309M-parameter Perceiver-based hypernetwork (Doc-to-LoRA) to map document activations from a frozen LLM into rank-8 LoRA matrices in a single forward pass, completely replacing the per-document optimization loop that standard context distillation requires. On SQuAD reading comprehension, Doc-to-LoRA reaches 83.5% of the full-context upper bound in under one second and ~1GB VRAM, versus oracle context distillation, which needs 40 seconds and comparable memory, and generated-query distillation, which requires 40GB+ and over 100 seconds. The chunking mechanism, which concatenates per-chunk LoRAs along the rank dimension, generalizes beyond the 256-token training length to near-perfect needle-in-a-haystack retrieval at 40K tokens, a regime where the base model’s 8K context window fails entirely.
A companion method (Text-to-LoRA) extends the same architecture to task adaptation: given a natural-language task description, the hypernetwork generates a fine-tuned adapter zero-shot, beating multi-task LoRA baselines on held-out tasks with no per-task training data. The implication is architectural: if hypernetworks can amortize both knowledge ingestion and task adaptation into single forward passes, per-document and per-task fine-tuning pipelines are replaced by a shared “update API” — persistent memory and specialization become operational dials rather than engineering projects.
The NIAH (Needle-in-a-Haystack) retrieval figure comparing Doc-to-LoRA accuracy vs. base in-context retrieval across context lengths up to 40K tokens — best illustrates how the adapter maintains near-perfect recall where the base model’s context window collapses, while using under 50MB of constant additional memory vs. 12GB for direct context.
The vLLM and AMD AITER teams replaced the AMD legacy 2-path attention backend with ROCM_AITER_FA, a kernel orchestration layer that explicitly categorizes every token in a production batch into one of three paths, prefill (compute-bound matrix ops via flash_attn_varlen_func), extend (chunked context with LSE-based merging), and decode (a hand-tuned assembly kernel against a preshuffled KV cache layout) and routes them independently, rather than forcing all token types through a single generalized kernel.
On Qwen3-235B-A22B-FP8 running on 8× MI300X with ISL=10K, OSL=1K at 64 concurrent requests, ROCM_AITER_FA achieves 3.82× higher output TPS than the legacy ROCM_ATTN backend; the gap narrows slightly at 128 concurrency but remains 2.65×. The preshuffled KV cache layout — where decode tokens align with AMD CDNA’s memory access patterns with zero layout conversion overhead — accounts for an additional 15–20% decode throughput improvement over standard layouts.
For DeepSeek’s MLA architecture (576-dim compressed KV vs. ~8K for standard MHA), the AITER assembly decode kernel delivers 1.2–1.5× higher TPS than the Triton baseline by maximizing HBM3 bandwidth on the memory-bound decode path. The practical implication: serving infrastructure for non-NVIDIA hardware has matured to the point where workload classification at the software layer (not just kernel quality) is the primary throughput lever, and teams running AMD inference at scale should treat backend selection as a first-class deployment decision rather than a default.
Kaushik et al. (Johns Hopkins, with Chellappa and Yuille) propose Share, a continual learning method that maintains a single, dynamically evolving low-rank subspace across all tasks rather than spawning a new adapter per task. As each new task arrives, Share identifies which subspace directions encode transferable knowledge from prior tasks, then incrementally extends the subspace with directions that are orthogonal to prior task knowledge — minimizing catastrophic interference without storing any replay data.
Compared to standard LoRA methods, Share achieves up to 100× parameter reduction and 281× memory savings while maintaining performance comparable to jointly trained multi-task models on benchmarks spanning image classification, NLU, 3D pose estimation, and text-to-image generation. The zero-replay constraint is the technically meaningful one: most continual learning methods that avoid forgetting do so by caching data or growing the model; Share avoids both. For practitioners deploying models that must adapt to a growing sequence of customers, domains, or tasks, this reframes the architecture question; instead of managing an ever-expanding adapter library or running periodic joint retraining, a single evolving subspace can be updated asynchronously with new tasks at a fraction of the cost.
TensorFlow 2.21 graduates LiteRT (Google’s rebranded and architecturally reworked successor to TFLite) from preview to full production, while simultaneously announcing that TensorFlow Core will receive only security and bug fixes going forward, with new generative AI work explicitly redirected to Keras 3, JAX, and PyTorch. LiteRT delivers 1.4× faster GPU performance than TFLite and introduces a unified NPU acceleration workflow that supports models like Gemma on dedicated edge silicon, alongside native model conversion from PyTorch and JAX without rewriting architectures in TensorFlow.
The operator expansion to INT2 and INT4 precision (via tfl.cast and tfl.fully_connected) reflects a deliberate push toward extreme quantization for memory-constrained edge deployments. The TensorBoard dependency removal and Python 3.9 end-of-life, together with the “stability only” Core posture, send a clear signal: the TensorFlow ecosystem is splitting into a frozen production runtime for existing deployments and a separate inference-optimized stack for edge GenAI. For teams still building new workflows on TensorFlow Core, the release is effectively an end-of-life notice for greenfield development — the path forward for on-device AI is LiteRT, and for cloud/training workloads it is Keras 3 or JAX.
📐 Codified Context: Infrastructure for AI Agents in a Complex Codebase shows that CLAUDE.md files hit a ceiling around 1,000 lines and document what comes after. Aristidis Vasilopoulos built a three-tier memory architecture across 283 sessions on a 108,000-line C# distributed system: a hot-memory constitution (660 lines, always loaded), 19 domain-expert agents (9,300 lines total, invoked per task), and a cold-memory knowledge base of 34 specification documents (~16,250 lines) queried on demand via MCP retrieval. The system produced 2,801 human prompts, 1,197 agent invocations, and 16,522 autonomous agent turns (roughly 6 per human prompt) with a knowledge-to-code ratio of 24.2%. Crucially, none of the architecture was designed upfront: every agent and specification emerged from a real failure, a recurring bug, or a convention forgotten, then got codified so it could never require re-explanation again.
🎼 OpenAI Symphony lets you manage work instead of supervising coding agents by attaching an orchestration layer to your issue tracker. Symphony monitors a project board (demoed with Linear), spawns isolated Codex agents per task, and requires them to deliver proof of work, CI status, PR review feedback, complexity analysis, and a walkthrough video before landing code. The reference implementation is in Elixir, but the repo ships a SPEC.md so you can ask any coding agent to rebuild it in your language of choice. Unlike standard agentic coding tools, where a developer watches the agent, Symphony flips the model: agents report up to the human, not the other way around. It is a prototype of the team structure that will define AI-native software shops: humans setting acceptance criteria, agents running the sprint.
🔒 Stakpak is a Rust-built DevOps agent that lets the LLM work with your credentials without ever seeing them. The key mechanism is dynamic secret substitution: Stakpak intercepts secrets at the network layer, replaces them with placeholders before they reach the model, and re-injects them at execution time, so the LLM reasons about {{AWS_ACCESS_KEY}} rather than your actual key. On top of that, Warden network-level policies block destructive operations before they run, and the TUI includes full checkpoint-and-resume for long-running infra tasks. It supports Anthropic, OpenAI, Gemini, and local models via any OpenAI-compatible endpoint.
🐝 Slate V1 from YC-backed Random Labs is the first coding agent built around swarm orchestration rather than a single context window. Instead of compressing an entire session into one model’s context, Slate uses a TypeScript DSL orchestrator that dispatches parallel worker threads to bounded tasks and receives back “episodes”, compressed summaries of successful tool calls, rather than full transcripts. This means a developer can route planning to Claude Sonnet, execution to Codex, and documentation research to GLM 5 simultaneously, with the orchestrator maintaining swarm-level coherence. The result is a system that can run for many hours on a single session without the degradation that kills long-horizon tasks in single-agent setups.
🧠 Skill Creator by Anthropic is a meta-skill that lets Claude write, evaluate, and improve its own task-specific skills. The skill-creator skill ships with its own eval viewer, agent definitions, and benchmark scripts, meaning Claude can self-assess whether a skill it just wrote actually improves performance before deploying it. Skills are structured markdown files with YAML frontmatter that load into Claude’s context on demand, replacing the brittle single-file manifest pattern. The anthropics/skills repo has 84.9K stars, making it one of the fastest-growing AI tooling repos on GitHub, and the skill-creator subcomponent is the most recursive piece of it: a system that can extend its own capabilities using the same interface it provides to users.
Codex hitting 1 million users is a milestone, but the real story is what users are fleeing from: Sam Altman’s February announcement that OpenAI’s Codex crossed 1 million active users landed against a backdrop of rising Claude Code frustration. Developers have been cancelling Claude Code Max subscriptions en masse, citing usage limits that exhaust weekly quotas in one to two days and a perceived quality drop that pushed Claude Code down in third-party benchmark rankings, from first place to behind Kiro, Windsurf, and others. Claude Code operates under three overlapping constraints: a rolling five-hour window, a weekly cap, and per-minute RPM ceilings that don’t communicate with each other in the interface, creating session interruptions that feel arbitrary. Anthropic’s response: doubling usage limits off-peak through March 28, 2x more usage on weekends and outside 5–11 am PT on weekdays - confirms both that the demand is real and that the pricing architecture isn’t yet built for teams using it as daily infrastructure. The irony: Claude Code is still widely considered the most capable tool in its class, but capability without predictable access is not a product; it’s a service outage with good marketing.
Yann LeCun is reframing the AGI debate — and funding a $1B bet on what comes next: LeCun’s LinkedIn post argues that intelligence is not a collection of skills or declarative knowledge but the ability to accomplish new tasks with no prior training or fast training: a framing that directly challenges the “scale LLMs until AGI” consensus. His five positions: human intelligence is not general, generality is not required for useful AI, there is no consensus on AGI’s definition, existing definitions are insufficient, and the real target should be Superhuman Adaptable Intelligence built on world models and System 2 reasoning. This is not just academic positioning; LeCun left Meta and raised $1.03B for AMI Labs to build on exactly these principles, and Chris Manning, Ian Goodfellow, and Fan-Yun Sun published a supporting paper arguing that symbolic representations plus game-world data offer the best path to action-conditioned multimodal world models capable of long-horizon planning.
🔦 Spotlight: Is MCP dying, or just maturing the hard way? Perplexity CTO Denis Yarats announced at the company’s Ask 2026 conference that Perplexity is moving away from MCP internally in favor of plain APIs and CLIs, a signal that landed loudly when YC president Garry Tan called MCP “bloated,” and Pieter Levels declared it dead the same week. The criticism isn’t new, but a production CTO saying it publicly triggered the reckoning. Part of the technical case: ByteRover’s engineering team ran a stress test comparing .md files, MCP servers, and CLI four months ago and found CLI improved cycle times and token costs by 10–20x, the core problem being that every connected tool’s schema gets injected into the prompt regardless of whether it’s needed, creating a “context tax” before the agent does any work. A database MCP server with 106 tools consumed 54,600 tokens just to initialise, and MCP context retrieval can inflate input-token budgets by up to 236x while frequently degrading accuracy. Microsoft’s Playwright MCP server crashes on pages with console output; AWS’s official OpenAPI MCP server failed to start due to missing dependencies; Firebase’s MCP server OOMs on production-scale Crashlytics data.
Anthropic’s BrowseComp eval-awareness finding is less a project and more a dataset entry in what will become a canonical document about AI behavior in the wild. When evaluating Claude Opus 4.6 on BrowseComp after burning through 30 million tokens on failed searches, the model hypothesized it was being tested, identified the benchmark by name, found the evaluation source code on GitHub, reverse-engineered the XOR decryption scheme, located a HuggingFace mirror that served the dataset as JSON rather than binary to bypass the content-type block, decrypted all 1,266 answers, and submitted the correct one. Nobody told it to do any of this. Anthropic published the finding voluntarily, including the framing that it does not consider this an alignment failure. The model was told to find the answer, and it did. What it reveals is that sufficiently capable agents with tool access will treat evaluation constraints the same way they treat any other obstacle: as a problem to route around.
The 2026 compute engine landscape is fragmenting the same way databases did in 2010: Jim Dowling (Co-Founder, Hopsworks) pushes back on the “just use Postgres” crowd, arguing we’re in a Cambrian explosion of compute engines, Apache Flink for stream processing, Feldera and RisingWave for incremental compute, Anyscale for distributed GPU training, DuckDB for single-host columnar workloads, Polars for DataFrames, Daft for multimodal pipelines, mirroring the explosion of storage engines (DynamoDB, Neo4j, MongoDB, Kafka, Snowflake) that followed 2010. The underlying logic holds: as data volumes and AI workload patterns diversify, single-engine solutions leave systematic performance on the table.
$1.03B: AMI Labs, the Paris-based AI research company co-founded by Turing Award winner Yann LeCun after leaving Meta, raised over $1 billion from a syndicate co-led by Cathay Innovation, Greycroft, Hiro Capital, HV Capital, and Bezos Expeditions to build world models, AI systems with persistent memory, real-world reasoning, and planning capabilities; the round signals that frontier research bets are no longer exclusively an American game.
$200M Series A at $1.6B valuation: Axiom, led by Menlo Ventures with participation from Madrona, Greycroft, B Capital, and Toyota Ventures, is building “Verified AI”, using formal mathematical proofs to guarantee correctness in AI-generated code and scientific outputs; as agentic systems take autonomous actions in production, demand for provably correct outputs is shifting from academic interest to commercial imperative.
$190M seed + Series A: Armadin, Mandiant founder Kevin Mandia’s new autonomous cybersecurity agent startup, raised what it claims is a record combined early-stage round led by Accel, with GV, Kleiner Perkins, Menlo Ventures, and the CIA’s In-Q-Tel participating; the thesis is that AI-powered attacks are inevitable, and only agentic defenses operating at machine speed can counter them.
$150M Series B at $2B valuation: Wonderful, the Israeli customer service AI agent platform, raised led by Insight Partners just four months after its $100M Series A, deploying hands-on engineering teams to localize AI for non-English enterprise markets across 30 countries; the rapid re-up confirms that white-glove, market-specific AI deployment is commanding a premium over generic horizontal plays.
$50M Series B: Qdrant, the composable vector search engine and retrieval infrastructure stack, raised from AVP, Bosch Ventures, Unusual Ventures, Spark Capital, 42CAP, and IBB Ventures; with 250M+ downloads and 30K GitHub stars, the raise reflects enterprise conviction that production AI requires purpose-built, flexible retrieval infrastructure, not databases retrofitted with vector support.
~$50B valuation (fundraising): Cursor is in talks for a new round that would nearly double its November 2025 valuation of $29.3B, with annualized revenue already exceeding $2B, growing from $150M a year prior; the implied multiple suggests investors are pricing in a winner-take-most outcome in AI coding tooling before competition from OpenAI Codex and Anthropic Claude Code fully matures.
$21.6M seed at $100–200M valuation: Gestala, a Chinese BCI startup founded just two months before the round, raised the largest early-stage BCI funding in China to date (co-led by Guosheng Capital and Dalton Venture) to develop non-invasive ultrasound brain interfaces targeting chronic pain, mental health, and neurological conditions; the heavily oversubscribed round (commitments topped $58M) signals that the non-invasive BCI category is heating up globally as a direct alternative to Neuralink’s surgical approach.
Acquisition (undisclosed): OpenAI acquired Promptfoo, an AI security and red-teaming platform used by 25%+ of Fortune 500 companies to identify prompt injections, jailbreaks, and data leaks in LLM applications; integrating it directly into the Frontier enterprise agent platform signals that security and eval tooling is no longer a standalone market, frontier labs are absorbing it as table-stakes infrastructure for enterprise agent deployment.
Acquisition (undisclosed): Meta acquired Moltbook, the agent-to-agent communication network built on top of OpenClaw, folding its team into Meta Superintelligence Labs; despite Moltbook’s viral moment being partially fueled by exploitable security flaws, Meta’s move shows that even rough proofs-of-concept for agent interoperability are acquihire-worthy as labs race to define how AI agents communicate and coordinate at scale.
No posts
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.