AI Briefing: 2026-06-25
Coverage window: 2026-06-23 – 2026-06-25 (last 48 hours; arXiv = 7-day window) Generated: 2026-06-25 00:05 UTC For: AI media professionals Sister publications: Stark News (crypto, daily)
🚨 Breaking (last 24h)
OpenAI unveils "Jalapeño" — first custom inference chip, built with Broadcom at gigawatt scale (Jun 24) — The biggest infrastructure story of the year so far. OpenAI's first "Intelligence Processor," designed from a blank slate for LLM inference, co-developed with Broadcom and Celestica in just nine months — what OpenAI calls the fastest ASIC development cycle ever in high-performance semiconductors. Engineering samples are already running ML workloads in lab at production target frequency/power, including GPT-5.3-Codex-Spark. Greg Brockman (President): "The world is moving to a compute-powered economy. Jalapeño is part of our long-term full-stack infrastructure strategy to make compute more abundant." Hock Tan (Broadcom CEO): "This is just the beginning of a multi-generation roadmap. By co-developing our industry-leading silicon directly with OpenAI, we are enabling the deployment of gigawatt scale data centers with Microsoft and other partners beginning in 2026." The chip is designed with flexibility to run all LLMs, not just OpenAI's. The same OpenAI models served to users helped accelerate parts of the chip design and optimization. Pairs directly with Jony Ive's
iohardware team to mark the moment OpenAI becomes a true vertically-integrated compute company. Sources: OpenAI: Jalapeño announcement · TechCrunch: OpenAI unveils first custom chip · Greg Brockman on OpenAI's in-house podcast · Hacker News discussion.Anthropic locks NSA out of Mythos amid escalating dispute (Jun 23–24) — Per the New York Times, the NSA lost access to Claude Mythos amid Anthropic's dispute with the US Commerce Department over the June 12 export-control directive that suspended Fable 5 and Mythos 5 access. First reported major downstream consequence of the export-suspension standoff. Signal that Anthropic is willing to absorb friction with federal intelligence customers to defend its export-policy posture. Source: NYT: NSA lost access to Mythos amid Anthropic dispute · Hacker News front page · Anthropic: Statement on US government directive (Jun 12).
Krea 2: 12B open-weights image model reaches SOTA for creative exploration (Jun 23) — Krea released Krea 2, a 12B-parameter open-weights diffusion transformer that places in the top 10 on the Artificial Analysis leaderboard for text-to-image and 2nd among independent labs. Architecture: DiT with iREPA, improved VAEs, Qwen3-VL encoder, grouped-query attention, sigmoid-gated attention, lightweight timestep modulation, and multilayer feature aggregation. Full multi-stage pipeline: pretraining → midtraining → SFT → preference optimization → RL. Notably: no AI-generated images in the pretraining mix — Krea argues even small proportions of synthetic data impose an upper bound on quality. Two control systems specifically for creative exploration: a prompt expander (SFT + RL on top of open-source LLMs to convert short prompts into richer visual directions) and a style-reference system for mixing the mood of multiple reference images. Source: Krea 2 Technical Report · Hacker News front page.
📊 Market Moves (last 48h)
Groq raises $650M to scale AI inference cloud — explicitly positioning against Nvidia (Jun 22) — Groq confirmed a $650M growth round led by Disruptive (Alex Davis, who is also Groq's chairman) and Infinitum. Funds scale toward 200 MW by end of 2027. Already 13 data centers across NA, Europe, ME, APAC. 5M+ developers, trillions of tokens/week. Critical context: this raise comes 6 months after Nvidia's non-exclusive licensing deal that hired away founder/CEO Jonathan Ross, president Sunny Madra, and other engineers — and Nvidia launched its own LPX inference hardware system at GTC 2026 using Groq's IP. Groq's thesis: "Inference will demand an estimated 15 to 20 times more compute" than training, and most AI clouds are built for training. New leadership: COO Alan Rice (ex-xAI/SpaceXAI, Meta Datacenters, US Navy nuclear submarine ops), CTO Sinclair Schuller (Apprenda, Nuvalence/EY), CPO Rakesh Malhotra (Microsoft cloud). Sources: Groq newsroom: $650M raise · TechCrunch: Groq confirms $650M raise.
Cerebras stock plunges after earnings as CEO clarifies margin guidance (Jun 24) — Cerebras shares dropped sharply after the company reported quarterly earnings; CEO Field clarified that earlier margin guidance had been "misunderstood" by the market. Continues the broader pattern of AI compute stocks repricing on investor scrutiny of unit economics. Source: TechCrunch: Cerebras stock plunges after earnings.
Anthropic continues Google brain drain — John Jumper leads wave of departures (Jun 20) — Nobel laureate John Jumper (Demis Hassabis's co-lead on AlphaFold 2) is leaving Google DeepMind to join rival Anthropic, per TechCrunch. Jumper shared the 2024 Nobel Prize in Chemistry for protein structure prediction. Part of an ongoing wave of Google AI researchers leaving for Anthropic, OpenAI, and other frontier labs. Source: TechCrunch: John Jumper leaving DeepMind for Anthropic · TechCrunch: AI researchers continue to leave Google for its rivals (Jun 24).
Anthropic SDK v0.112.0 (Jun 24) — New release with support for
system.messagestreaming events, a memory-tool fix for parent-directory permissions, a new refusal category, andUser Profile IDheader support. Source: GitHub: anthropic-sdk-python v0.112.0.OpenAI Python SDK v2.44.0 (Jun 24) — Bug fix: auth header priority. Source: GitHub: openai-python v2.44.0.
🔬 Research (7-day arXiv window)
Top papers (latest batch: 2026-06-23):
[2606.24597] Qwen-AgentWorld: Language World Models for General Agents (Jun 23, Alibaba Qwen team) — First language world models capable of simulating agentic environments across 7 domains via long chain-of-thought reasoning. Two variants: Qwen-AgentWorld-35B-A3B and Qwen-AgentWorld-397B-A17B. Three-stage training: CPT (continual pretraining on state-transition dynamics + augmented corpora) → SFT (activates next-state-prediction reasoning) → RL (sharpens simulation fidelity via hybrid rubric-and-rule rewards). New benchmark: AgentWorldBench, built from real-world interactions of 5 frontier models on 9 established benchmarks. Demonstrates the framework is also useful as a decoupled environment simulator (scalable RL training that surpasses real-environment training alone) and as a warm-up objective for unified agent foundation models (improves downstream performance across 7 agentic benchmarks). arXiv · Code.
[2606.24855] OpenThoughts-Agent: Data Recipes for Agentic Models (Jun 23) — Public recipes for curating training data for broadly capable agents. Compares against SWE-Smith and other open efforts. arXiv.
[2606.24884] InSight: Self-Guided Skill Acquisition via Steerable VLAs (Jun 23) — Vision-language-action framework that lets VLA models learn manipulation skills beyond those in their training data. arXiv.
[2606.24842] World Models in Pieces: Structural Certification for General Agents (Jun 23) — Argues that in the big-world regime, agents cannot be universally capable; introduces structural certification framework. arXiv.
[2606.24849] IV-CoT: Implicit Visual Chain-of-Thought for Structure-Aware Text-to-Image Generation (Jun 23) — Unified MLLM technique for structure-aware prompt following (object counts, spatial relations). arXiv.
[2606.24874] FLUX3D: High-Fidelity 3D Gaussian Generation with Diffusion-Aligned Sparse Representation (Jun 23) — New 3DGS generation method that preserves high-frequency details. arXiv.
[2606.24878] AIR: Adaptive Interleaved Reasoning with Code in MLLMs (Jun 23, recent batch) — Interleaved code+reasoning for multimodal LLMs. arXiv.
🛠️ Tools (last 48h)
OpenClaw v2026.6.10 STABLE (Jun 24, 5,182-char changelog) — Headline features:
- Automatic fast mode for talks — OpenClaw enables fast mode for short conversational turns, then returns to normal mode for longer runs with bounded fallback. (#85104)
- More reliable model routing — Zai model synthesis, GLM overload failover, native reasoning-level selection now follow the active model catalog more consistently. (#94461, #93241, #94067, #94136)
- Safer session and channel state — channel switches reset stale origin fields; cron delivery awareness stays attached to the target session. (#95328, #93580)
- Trusted policies survive hook composition — composed hook registries keep the trusted tool policies required by approval-sensitive flows. (#94545)
- 12 merged PRs; full release evidence published. Source: GitHub: openclaw v2026.6.10 · npm package · Release SHA: aa69b12d0086b631b139c1435c9621a5783e3a40.
OpenClaw v2026.6.11-beta.1 (Jun 24, 37,558-char changelog, 305 merged PRs) — Substantial infrastructure-level release:
- Slack relay mode for incoming messages — open Slack channels as agent ingress surfaces. (#94707)
- Mattermost
/oc_queuenative slash command — slash-command level integration with Mattermost. (#95546) - Per-DM model overrides — assign different models to different DM threads for cost/perf tuning. (#95120)
openclaw agent --message-file— practical file-driven agent invocations. (#93351)- RAFT CLI wake bridge — connect OpenClaw to the Raft external agent network. (#95497)
- Bundled plugin icon metadata — official plugins externalized cleanly with icon support. (#95683, #95845)
- Android settings detail panels — better mobile configuration visibility. (#95148)
- Codex partial deltas, harness activation, long-context prompt-cache stability — reliability fixes for Codex sessions. (#95404, #95652, #95624)
- Provider/model coverage expansion — catalog parsing, reasoning controls, provider model resolution, encrypted reasoning support for more live providers. (#95283, #95710, #95268, #95744, #95686, #93956)
- Channel delivery fixes — Telegram progress rendering, webhook lifecycle, reaction directives, duplicate mirror writes, queued update draining, WhatsApp durable reply targets. (#95532, #93002, #95183, #94506, #94977, #95069, #95577, #95007, #95914)
- Source: GitHub: openclaw v2026.6.11-beta.1.
💭 Industry Pulse (last 48h)
"In the Weights" launches as AI-centric vanity search (Jun 20–24, trending) — Ex-OpenAI designers Thomas Dimson and Joey Flynn (both joined OpenAI via the Global Illumination acquisition) launched In the Weights, a website that scores how well LLMs recall a given person without web search. Queries Grok, Gemini, multiple GPT versions, Claude, Llama, and other models. Assigns a "strength score" — currently Macaulay Culkin at 988, Luciano Pavarotti close behind. Quote (Dimson): "Being in the weights means your existence was deemed important in the process of creating superhuman artificial intelligence." Quote (Dimson on why he built it): "Google vanity searches are the wrong objective in 2026 as more traffic moves to LLMs." Source: TechCrunch: In the Weights.
Companies scramble to stop employees from maxing out AI budgets with small tasks (Jun 24) — TechCrunch reports a growing pattern: enterprise AI budgets being exhausted not by single big projects, but by large numbers of low-cost quick tasks. The economics of inference-heavy workloads (every chat, every quick code completion, every small research query) mean that even sub-$0.01 calls add up at company scale. Per the TechCrunch piece, internal AI governance teams are racing to add hard caps, per-user rate limits, and prompt-level cost controls. Source: TechCrunch: Companies scrambling to stop AI budget maxing.
Engineering jobs prove most resilient to AI automation — counter to early predictions (Jun 24) — TechCrunch summary of new labor-market data showing engineering roles holding up far better than originally forecast when generative AI exploded in 2023. Pairs with continued enterprise AI tool deployments (Samsung Electronics, LG, NAVER, etc.). Source: TechCrunch: AI was supposed to kill engineering jobs.
Agility Robotics plans to go public via SPAC at $2.5B (Jun 24) — Humanoid robotics company Agility Robotics announces SPAC merger at $2.5B valuation. Continues 2026 humanoid-robotics IPO wave. Source: TechCrunch: Agility Robotics SPAC.
🖼️ New Presentations
No presentations generated this run. Two presentation triggers were detected (OpenClaw v2026.6.10 stable + v2026.6.11-beta.1 with 305 PRs, and Hermes Agent v2026.6.19 from Jun 19). Per the briefing iteration-budget policy, these have been logged in the relevant wiki entity pages and will be processed by the dedicated version-update-presentation-pipeline job.
📡 Sources & Data Provenance
| Source | Status | URL |
|---|---|---|
| GitHub API (releases) | ✅ | https://api.github.com |
| arXiv API (cs.AI/cs.LG/cs.CL) | ✅ | https://arxiv.org |
| OpenAI newsroom (via jina.ai) | ✅ | https://openai.com/news |
| Anthropic newsroom (via jina.ai) | ✅ | https://www.anthropic.com/news |
| TechCrunch AI category (via jina.ai) | ⚠️ Degraded | https://techcrunch.com/category/artificial-intelligence/ |
| Hacker News front page (Jun 24) | ✅ | https://news.ycombinator.com/front?day=2026-06-24 |
| Krea blog (via jina.ai) | ✅ | https://www.krea.ai/blog/krea-2-technical-report |
| Groq newsroom (via jina.ai) | ✅ | https://groq.com/newsroom |
| Twitter/X (twitter-api.io) | ❌ HTTP 401 Unauthorized | https://twitter-api.io |
| Web Search (DuckDuckGo/Python) | ⚠️ Skipped | https://duckduckgo.com |
| Wiki Raw Archive | ✅ Used as fallback | ~/wiki/raw/articles/ |
Twitter API status (Jun 25, 2026): Still returning HTTP 401 Unauthorized — confirmed against X_API_SECRET from ~/.hermes/.env. No commentary from monitored accounts (@karpathy, @sama, @ylecun, @gdb, @AndrewYNg) included in this briefing. If the API recovers, tomorrow's briefing will catch up on social commentary.
📚 Sources & References
OpenAI Jalapeño chip:
- OpenAI: Jalapeño announcement
- OpenAI: October 2025 Broadcom partnership announcement
- TechCrunch: OpenAI unveils first custom chip built by Broadcom
- Greg Brockman on OpenAI's in-house podcast
- Reuters: OpenAI custom chip rumors (Feb 2025)
- Google Cloud TPU · AWS Trainium — prior art for AI-accelerator custom chips
Anthropic / NSA Mythos dispute:
- NYT: NSA lost access to Mythos amid Anthropic dispute
- Anthropic: Statement on US government directive (Jun 12, 2026)
- Hacker News front page
Krea 2:
- Krea 2 Technical Report
- Hacker News front page
- Artificial Analysis leaderboard — Krea 2 in top 10 for text-to-image
Groq:
- Groq newsroom: $650M raise
- TechCrunch: Groq confirms $650M raise
- Nvidia LPX inference system
- TechCrunch: Groq reportedly raising $650M (May 29)
- Baseten $1.5B (Jun 18, per TechCrunch) — corroborating inference-market heat
Cerebras:
Anthropic talent:
- TechCrunch: John Jumper leaving DeepMind for Anthropic
- TechCrunch: AI researchers continue to leave Google for its rivals
SDK releases:
arXiv papers (7-day window):
- 2606.24597 — Qwen-AgentWorld · Code
- 2606.24855 — OpenThoughts-Agent
- 2606.24884 — InSight
- 2606.24842 — World Models in Pieces
- 2606.24849 — IV-CoT
- 2606.24874 — FLUX3D
OpenClaw releases:
Industry Pulse:
- In the Weights · TechCrunch coverage
- TechCrunch: Companies scrambling to stop AI budget maxing
- TechCrunch: AI was supposed to kill engineering jobs
- TechCrunch: Agility Robotics SPAC at $2.5B
Unlinked claims were cross-referenced from multiple sources. Twitter/X commentary from monitored accounts not included — API returning HTTP 401 as of Jun 25, 2026 (verified).