AI High Signal Digest

OpenAI unveiled Jalapeño, its first custom inference chip. Built with Broadcom for ChatGPT, Codex, the API, and future agentic products, it extends OpenAI’s stack into infrastructure @OpenAI. OpenAI said the program went from initial design to tape-out in nine months, used ChatGPT in the engineering process, is already running GPT-5.3-Codex-Spark in the lab, and targets substantially better performance per watt, with deployment planned for end-2026 @kimmonismus +1. The stated goal is lower dependence on external GPUs and tighter control over compute economics @kimmonismus.

Google launched computer use in Gemini 3.5 Flash. The feature lets an agent take a screen and a goal, then determine actions across browser, mobile, and desktop environments @_philschmid. Shared examples show it auditing docs pages by navigating, running code snippets, taking screenshots, and returning a report, with safeguards including user confirmation and auto-stop on prompt injection @_philschmid.

MirendilAI launched with a $200M seed round. The startup says it will build self-accelerating AI R&D systems, with a 20-person founding team from Anthropic, xAI, Google DeepMind, and OpenAI; the round was led by a16z and Kleiner Perkins, with a major NVIDIA investment @bneyshabur. Its pitch is to democratize frontier AI capabilities for broader scientific use rather than concentrate them in a few labs @bneyshabur.

GLM-5.2 posted the strongest ARC-AGI-2 result yet for an open-source model. Verified scores were 22.8% on ARC-AGI-2 at $0.25 and 77.0% on ARC-AGI-1 at $0.19, with ARC Prize saying performance is comparable to GPT-5.4 and GPT-5.5 at low reasoning effort @arcprize. François Chollet called it the strongest ARC-AGI-2 performance to date by an open-source model @fchollet.

Zyphra argued that continual learning hits a deeper failure mode than forgetting. Its new work identifies plasticity loss as models losing the ability to learn new data, shows it across 5M-314M parameter GPT-style models, and reports the same decline even in stationary pretraining @ZyphraAI +2. The team fit a scaling law for the onset, T ∝ P^0.83, suggesting scale delays the problem but does not remove it @ZyphraAI.

AI21 topped DeepResearch Bench II by merging weak agents instead of building a new one. It combined seven agents ranked 7-13 into a single report pipeline and reported a new #1 score of 64.38 @AI21Labs +1.

OpenAI updated GPT-5.5 Instant. The new version is described as better at understanding user intent, adapting responses, handling complex constraints, and improving shopping and local recommendations; rollout started today for paid users and tomorrow for free users @OpenAI.

Perplexity launched Computer for Counsel. The product connects legal research databases, document tools, and matter-management systems so lawyers can pull citable sources from tools including MidpageAI, LegalZoom, Docusign, and NetDocuments @perplexity_ai. It is available to Pro and Max subscribers @perplexity_ai.

Notion and Cursor pushed agents deeper into team workflows. Notion introduced External Agents with Claude and Cursor so teams can assign work from shared boards and @-mention agents like teammates @NotionHQ. Cursor said the integration runs on its SDK so cloud agents can take tasks from Notion and open PRs using the same runtime as Cursor itself @cursor_ai.

Qualcomm agreed to acquire Modular. Both sides said the deal is meant to unify accelerated compute with an open platform spanning edge to cloud and hardware from CPUs and GPUs to NPUs and custom ASICs @Qualcomm +2.

Anthropic is pulling more talent from Google DeepMind. Bloomberg reported that Jonas Adler and Alexander Pritzel, both viewed internally as key Gemini contributors, are leaving Google for Anthropic @EdLudlow.

The first legal challenge to the Trump administration’s AI export controls has arrived. Legion is suing over the forced shutdown of Anthropic’s Fable 5 and Mythos 5 for foreign nationals, arguing export-control laws do not cover access to hosted AI models or text outputs and that no national emergency was declared @kimmonismus +1. The case turns on whether hosted frontier-model access can be treated as export-controlled technology when users only receive text outputs @kimmonismus.

Europe joined a US-led AI supply-chain pact. The EU, Germany, the Netherlands, and Greece joined Pax Silica, covering chips, critical minerals, energy, and compute; Jacob Helberg explicitly positioned it against digital sovereignty built around duplicative national tech stacks @kimmonismus.

Anthropic launched Claude Tag, turning Claude into a Slack teammate. Claude can join selected channels with access to chosen tools, data, and codebases for async task delegation @claudeai +1. Anthropic says the Claude Code team has used it internally all year and that Claude now writes 65% of its product code; Tag is in beta for Claude Enterprise and Team plans @ClaudeDevs +1. Karpathy called this the “3rd major redesign of LLM UIUX,” centered on persistent, asynchronous agents with org-wide context @karpathy.

Speech AI moved closer to full conversational context. AssemblyAI launched Universal-3.5 Pro Realtime, which uses the agent side of a conversation as context for transcription @AssemblyAI. The company says one team cut error rates on critical utterances from 26% to 9% with that feature @AssemblyAI. At the same time, Artificial Analysis launched a Speech to Speech Index, with GPT-Realtime-2 leading overall and Grok Voice Think Fast 1.0 leading the agentic-performance subscore @ArtificialAnlys +1.

Cyber agencies shortened the AI risk timeline. The Five Eyes alliance warned organizations they have months, not years, to protect systems from accelerating cyber threats driven by frontier AI @NCSC.

Qwen-AgentWorld introduced language world models for agentic simulation. The release includes 35B-A3B and 397B-A17B models described as the first language world models able to simulate agentic environments across seven domains, with “Decouple” and “Unify” strategies for applying them to agents @iScienceLuvr.

ParallelKernelBench showed how far LLMs still are from reliable multi-GPU kernel generation. The benchmark covers 87 real problems from codebases including Megatron-LM, DeepSpeed, TensorRT-LLM, and NeMo-RL @asplencmnt +1. Best zero-shot performance reached 28/87 correct, while an agentic compile-test-profile-revise loop improved Gemini 3 Pro from 24 to 35/87@togethercompute +1.

DFlash pushed speculative decoding forward on Blackwell GPUs. NVIDIA says the open-source block diffusion drafter can raise inference throughput by up to 15x while maintaining responsiveness @NVIDIAAI. vLLM reported 4.4x–5.8x gains on Gemma-4 31B, with drop-in support via vLLM, SGLang, and TensorRT-LLM @NVIDIAAI +1.

Krea 2 open weights shipped in two forms: Krea 2 Raw for fine-tuning and Krea 2 Turbo as a faster distilled model with broad aesthetic range @krea_ai. Krea also published the code, weights, and technical report, while Ostris added day-0 LoRA support and reported strong early fine-tuning results on a hard “omniface” concept @fal +1.

Baidu open-sourced Unlimited OCR for long-document transcription. The model has 3B total parameters with 500M active, said it sets new SOTA on OmniDocBench v1.5/v1.6, and can transcribe 40+ pages in one forward pass using Reference Sliding Window Attention @BaiduAI_News.

NVIDIA launched the BioNeMo Agent Toolkit for workflows such as protein structure prediction, docking, generative chemistry, and genomics, with Baseten making all 10 BioNeMo NIMs available on day one @NVIDIAHealth +1.

OpenAI highlighted the scale of its recent developer-platform expansion. The company says it shipped 30+ API models, features, and upgraded tools in the last six months, including GPT-5.5, GPT-Realtime-2, GPT-Image-2, new agent-building blocks, the OpenAI CLI, and Bedrock availability @OpenAIDevs.

The UK consolidated five AI labs into the new BOLD Lab. BOLD says it is focused on beyond-backprop methods, human-centric learning, and embodied learning, with £30M in backing from UKRI and EPSRC @bold_lab_ai.

Together AI pointed to a new scale marker for open-model production use. The company said 400T tokens now reflects real workload adoption, driven by frontier-quality open models, better token economics, and more control over inference @togethercompute.

OpenAI shifted cyber AI from detection toward remediation. Daybreak now includes GPT-5.5-Cyber, Codex Security, a Cyber Partner Program, and Patch the Planet; OpenAI says the system can find and generate patches for flaws across major browsers, network infrastructure, operating systems, and widely used open-source projects. Since March, it says 30M+ commits have been scanned and 70K+ findings marked fixed @OpenAI +2.

GLM-5.2 is giving open weights a stronger claim on real work. Artificial Analysis ranked it #3 overall on GDPval-AA at 1524 Elo and the top open-weights model by a wide margin; on AA-Briefcase, GLM 5.2 sits within 90 Elo of Claude Opus 4.8 at $2.40 per task, or 65% lower cost@ArtificialAnlys +1.

AI compute demand is showing up as rented cluster capacity at extreme scale. SpaceX's Colossus clusters are now tied to $2.32B in monthly deals across Anthropic, Google, and Reflection, with all three structured as short-term agreements carrying 90-day out clauses @jaminball.

A large audit challenged common LLM-as-a-judge metrics. Across roughly 541,000 judgments from 21 judges, researchers found exact-match agreement overstated skill; switching to Cohen's kappa cut agreement by 33-41 points on MT-Bench and moved rankings by up to 14 places@dair_ai.

TMax made agentic RL more reproducible. The release includes open terminal-agent models plus data, weights, and rollouts; the team says a standard training job used 8 H100 nodes for 2-3 days, and getting the recipe right took O(100) jobs @hamishivi +1.

A reasoning-style distillation improved local orchestration. A LoRA distillation of DeepSeek V4 Pro traces into Qwen3.6-35B-A3B raised GPQA-Diamond from 72.7 to 80.3 and cut average agent orchestration time from 60.7s to 26.6s@ZhihuFrontier.

Google's Interactions API is now GA. Google says it is the primary interface for Gemini models and agents, with one API for models and agents, background execution, multimodal generation, and an isolated Linux sandbox via Antigravity Agent @Google +1.

GitHub Copilot added Agent merge. The feature lets an agent create a PR, run actions, do code review, and prepare the merge; early users described it as a major improvement in getting agent-written PRs over the finish line @JamesMontemagno +1.

Delos launched persistent AI workers. Workers keep identity and memory across tasks, get their own email, phone number, and Slack handle, and Delos says the launch reached $1M ARR in a couple of days @pierre_dlgr +1.

Baseten raised $1.5B to expand inference infrastructure. The company says it is building the Inference Cloud so customers can run AI products with speed, reliability, and control as more teams shift toward open and specialized models @baseten.

Micron and Anthropic tied frontier models to the hardware stack. Their strategic agreement spans memory and storage AI architecture design, supply, enterprise Claude adoption inside Micron, and a strategic Anthropic investment @firstadopter.

OpenAI pushed health AI on both product and research fronts. GPT-5.5 Instant is now on par with OpenAI’s frontier Thinking models for health questions, with better urgent-care detection, context gathering, and uncertainty communication for the 230M+ weekly health queries ChatGPT sees; possible factuality errors fell 71%, and the model is free to all users @OpenAI +1. In parallel, OpenAI, Boston Children’s Hospital, and Harvard reported in NEJM AI that o3 Deep Research helped clinicians find 18 diagnoses across 376 previously unsolved pediatric cases, with every result undergoing human adjudication @OpenAI +2.

New agent benchmarks were a reality check for long-horizon work. AA-Briefcase evaluates multi-week projects with thousands of messy inputs, including documents, transcripts, 25,000+ Slack messages, and 3,500+ emails @ArtificialAnlys. Claude Fable 5 leads at 1587 Elo, but it satisfies all rubric criteria on only 3% of tasks, and no model clears 50% on 31 of 91 tasks @ArtificialAnlys. Terminal-Bench Challenges reported a similar pattern: even the strongest frontier models still score very low on large-scale autonomous software tasks @terminalbench +1.

GLM-5.2 kept strengthening the case for open models. It is now the top open model on Agent Arena at #10 overall @arena +1, scored 1266 Elo on AA-Briefcase at an average cost of $2.40 per task @ArtificialAnlys, and can now run locally in a 2-bit version that shrinks from 1.51TB to 238GB while retaining about 82% accuracy @UnslothAI. The notable shift is that the story is no longer just leaderboard strength; it is also price and local execution.

OpenAI released new work on broadly beneficial RL. Using reinforcement learning on realistic conversations across 12 domains, the trained model improved on 44 of 53 independent evaluations spanning deception, reward hacking, safety, health, and mental health @OpenAI +2. Health-only training also improved non-health misalignment, deception, and reward-hacking evaluations, and the model was harder to steer toward harmful behavior with adversarial prompts @OpenAI +1.

Sakana AI introduced Doc-to-LoRA and Text-to-LoRA. The methods use a hypernetwork to generate LoRA adapters on demand, letting models specialize to new tasks or internalize documents with sub-second latency @SakanaAILabs. In experiments, Doc-to-LoRA reached near-perfect needle-in-a-haystack accuracy on inputs five times longer than the base model’s context window and could transfer visual information from a vision-language model into a text-only LLM @SakanaAILabs.

Perplexity launched Brain in Computer, a continuously learning memory system that builds a context graph from sessions, files, and connectors; on context-heavy tasks it improved answer correctness by 25%, recall by 16%, and ran 13% cheaper per task @perplexity_ai +2.

Claude Code added Artifacts, interactive pages built from a session, such as PR walkthroughs or living dashboards, shared through private team links on Team and Enterprise plans @claudeai +1.

OpenAI added Codex Record & Replay, which turns a demonstrated recurring workflow into an inspectable, editable skill; recording is user-controlled and the rollout starts in select markets @OpenAIDevs +1.

OpenAI hired Dean Ball to lead a new Strategic Futures team focused on shaping frontier AI policy, starting July 6 @deanwball.

Poolside paired a model release with a clearer strategy signal. It released Laguna M.1 under Apache 2.0 and said open weights are now its default @poolsideai +1.

Taste Labs emerged from stealth with an $18.5M seed. Its pitch is building the data and infrastructure layer that gives models and agents taste, and it says it is already working with frontier labs on post-training data and RL environments @thaiscbranco_.

The White House and Anthropic are developing a formal jailbreak-severity framework, with proposed benchmarks for how much safeguards were bypassed, what capabilities were exposed, and the practical consequences of a breach @SophiaCai99.

Google DeepMind published its AI Control Roadmap for managing advanced AI systems inside Google, arguing most agent failures come from misinterpreting commands or over-pursuing goals, and warning there is a narrow window to embed structural security protocols before multi-agent systems scale @GoogleDeepMind +2.

GLM-5.2 broke out as a major open-model release. Z.ai released GLM-5.2 with MIT-licensed open weights, major coding and agentic gains, a 1M-token context window, dual reasoning modes, and unchanged API pricing @Zai_org. Within hours, it ranked #1 on Design Arena, #10 overall on Agent Arena as the top open model, and #2 in Code Arena: Frontend @Designarena +2. That makes it one of the strongest open models now showing up across both coding and long-horizon agent leaderboards @arena +1.

SpaceX moved to buy Cursor and said a joint model is already on the way. SpaceX said it exercised its option to acquire Cursor in an all-stock transaction, and the companies said they have already been jointly training a model that will be released in Cursor and Grok Build soon @SpaceX. Separate posts describing the merger agreement valued Cursor at $60B and pointed to a Q3 2026 close, while Cursor said users should expect significant improvements soon @xdNiBoR +1. This turns the deal into both a major acquisition and a near-term product integration story @SpaceX +1.