DISPATCH

Everything that mattered in AI, one page a week.

Most AI news does not survive the week. This is the part that did — the releases, the research, and the shifts that actually change how we build. Designed & built to keep you up to date with things in AI without needing to be unemployed. Just refresh Saturday morning and review the last week's dispatch.

90 DISPATCHESWRITTEN EVERY FRIDAYNEXT UPDATE IN —

DISPATCH 18

WEEK OF APR 26 – MAY 2, 2025

Qwen3 opened eight models under Apache 2.0 the same week the leaderboard itself got audited

A 235B mixture-of-experts activating 22B parameters landed under a license with no acceptable-use clause, beside a 671B open-weight Lean prover. Simultaneously, researchers documented how the arena rankings favored closed labs — and named 27 private Llama-4 variants.

Alibaba's Qwen team released Qwen3 on April 29: eight open-weight models under Apache 2.0, two of them mixture-of-experts — 235B total with 22B active, and 30B total with 3B active — plus dense checkpoints at 32B, 14B, 8B, 4B, 1.7B and 0.6B, each runnable in a thinking or non-thinking mode toggled per request. Activating 22B parameters per token carries roughly the memory and FLOP profile of a 20-30B dense model, so frontier-adjacent quality now deploys on hardware teams already own — and Apache 2.0 removes the acceptable-use and reporting clauses that turned some earlier open weights into a legal review.

The same week produced the correction that makes such claims checkable. The Leaderboard Illusion, a 68-page audit from Cohere Labs and academic collaborators, documented that a small group of preferred providers could test many variants privately and publish only the best-scoring one, identified 27 private Llama-4 variants Meta tested ahead of release, and simulated that ten private variants were worth roughly 100 Arena points. DeepSeek meanwhile published a 671B open-weight theorem prover for Lean 4, and Meta and Microsoft priced the compute build-out in opposite directions on the same day.

WED · Apr 30, 2025open-weightsdeepseekformal-methodstheorem-proving

DeepSeek publishes a 671B open-weight theorem prover for Lean 4

DeepSeek uploaded DeepSeek-Prover-V2 to Hugging Face on April 30: a 671B-parameter open-weight model built on DeepSeek-V3-Base that writes Lean 4 proofs, plus a 7B sibling with a 32K context window. The pipeline bootstraps itself — DeepSeek-V3 decomposes a theorem into Lean subgoals, a smaller prover searches each subgoal, and the resolved steps are stitched back into chain-of-thought data used to cold-start reinforcement learning. DeepSeek reports an 88.9% pass ratio on MiniF2F-test and 49 of 658 PutnamBench problems solved, which it calls state of the art for neural theorem proving.

Formal verification is the one part of the stack where generated code can be checked instead of trusted, and the missing piece was always a prover good enough to automate it. With 671B weights downloadable and a 7B variant that fits on a single GPU, a team can wire proof search into CI and let Lean reject bad proofs rather than asking a second model for an opinion. The marginal cost of that loop is the hardware you already own, not an API invoice.

DeepSeek publishes a 671B open-weight theorem prover for Lean 4 — Hugging Face

WHY IT MATTERS

If machine-checked proofs become a cheap local primitive, verified code moves from research demo to CI gate: an agent proposes a patch, Lean decides whether the reasoning holds, and reviewers stop being the last line of defense on correctness claims. That is a materially different reliability story from asking a second model to agree with the first.

Hugging Face: DeepSeek-Prover-V2-671B GitHub: DeepSeek-Prover-V2 (method, ProverBench, 7B variant)

WED · Apr 30, 2025infrastructurecapexearningsmetamicrosoft

Meta raises 2025 AI capex to $64-72B while Microsoft says its capex growth slows

Meta's Q1 2025 results on April 30 lifted full-year capital expenditure guidance to $64-72 billion from $60-65 billion, citing additional data-center investment for AI and a higher expected cost of infrastructure hardware; the same release puts Meta AI at almost 1 billion monthly actives. Microsoft reported the same day: revenue of $70.1 billion, Azure growth of 33%, and $21.4 billion of capex in the March quarter, up 53% year on year. CFO Amy Hood told the call that capex keeps growing in FY26 but at a lower rate than FY25, weighted toward short-lived assets correlated with revenue — from a company that had spent recent months slowing or pausing some data-center projects.

The two numbers price capability in opposite directions on the same day. A raised Meta budget says capacity is scarce enough to justify paying more for servers; Microsoft's decelerating curve and shift to revenue-linked hardware says the build-out is being sized to demand rather than ambition.

Meta raises 2025 AI capex to $64-72B while Microsoft says its capex growth slows — s21.q4cdn.com

WHY IT MATTERS

Hyperscaler capex guidance is the twelve-month leading indicator for GPU availability and inference pricing. A budget being raised on hardware costs while a competitor's growth rate falls tells you to model per-token costs as flat-to-uneven rather than falling, and to keep a self-hosted path priced alongside any API contract.

Meta: Q1 2025 results (capex guidance $64-72B)CFO Dive: Microsoft capex to grow at slower rate Tech Brew: Meta, Microsoft and Google stick to AI spending plans

TUE · Apr 29, 2025open-weightsqwenmoepricingdistillation

Qwen3 opens eight models under Apache 2.0, including a 235B MoE that activates 22B

Alibaba's Qwen team released Qwen3 on April 29: eight open-weight models under Apache 2.0, two of them mixture-of-experts — Qwen3-235B-A22B (235B total, 22B active, 128K context) and Qwen3-30B-A3B (30B total, 3B active, 128K) — plus dense 32B, 14B, 8B, 4B, 1.7B and 0.6B checkpoints. Every model runs in two modes, a thinking mode that emits a reasoning trace and a non-thinking mode for latency-critical calls, toggled per request through the chat template's enable_thinking flag. Qwen reports the flagship as competitive with o1, DeepSeek-R1, o3-mini, Grok-3 and Gemini 2.5-Pro on coding, maths and general benchmarks, and says the 30B MoE beats QwQ-32B while activating a tenth of the parameters.

The cost consequence is arithmetic: activating 22B parameters per token carries roughly the memory and FLOP profile of a 20-30B dense model, so frontier-adjacent quality now deploys on hardware teams already own. Supported checkpoints landed in vLLM and Ollama on day one alongside OpenAI-compatible endpoints, so moving a workload off a closed API is a config change rather than a migration. Apache 2.0 also removes the acceptable-use and reporting clauses that turned some earlier open weights into a legal review, and the small dense models exist mainly as distillation targets for teams that want the behavior at a fraction of the footprint.

WHY IT MATTERS

A permissively licensed, frontier-adjacent MoE resets the build-versus-buy line: prompts, weights and fine-tunes stay inside your infrastructure, inference becomes a hardware amortization line you control, and every closed-model price has to justify itself against a downloadable substitute that scores within a few points. This is the release that made self-hosting a reasoning-class model a mainstream default rather than a compromise.

Qwen: Qwen3 — Think Deeper, Act Faster Qwen3 model collection on Hugging Face (Apache 2.0)QwenLM/Qwen3 on GitHub

TUE · Apr 29, 2025evalsbenchmarkslmarenaopen-weightsmeta

The Leaderboard Illusion: 27 private Llama-4 variants, and an arena that favored closed labs

On April 29, researchers from Cohere Labs and academic collaborators posted The Leaderboard Illusion (arXiv:2504.20879, 68 pages), an audit of Chatbot Arena's ranking pipeline. They document an undisclosed practice that let a small group of preferred providers test many variants privately and publish only the best-scoring checkpoint, and at the extreme they identify 27 private LLM variants Meta tested in the lead-up to the Llama-4 release. On data, they estimate Google and OpenAI received 19.2% and 20.4% of all arena battles respectively, while 83 open-weight models combined received 29.7%. Simulations put the payoff of ten private variants at roughly 100 Arena points, and two identical checkpoints of the same model scored 17 points apart purely from sampling randomness.

This lands three weeks after Llama 4's own leaderboard problem, when Meta submitted a non-public experimental chat version of Maverick to LMArena and LMArena replied that Meta's reading of its policy did not match what it expects from model providers. The paper's contribution is that the anecdote generalizes: sampling rates, model deprecation and score retraction are all asymmetric, so single-digit rank gaps carry no signal.

WHY IT MATTERS

Model selection increasingly starts from a leaderboard number. A documented, provider-favoring distortion means that number is marketing input rather than evidence — and the same overfitting it rewards in labs is the failure mode you inherit if you route production traffic by rank instead of by eval on your own workload. The operational fix is cheap: evaluate on your own traffic and treat Arena rank as a sourcing hint.

arXiv: The Leaderboard Illusion Full HTML text

TUE · Apr 29, 2025agent-infrastructureguardrailsprompt-injectionllamasecurity

LlamaCon: an Llama API preview, a standalone Meta AI app, and open guardrails for agents

At its first LlamaCon on April 29, Meta opened a limited free preview of the Llama API with Python and TypeScript SDKs, OpenAI-SDK compatibility, and hosted fine-tuning plus an evaluation suite starting with Llama 3.3 8B — with the explicit promise that models built there remain portable to self-hosting. The same day it shipped the standalone Meta AI app, pitched on the personalization it can draw from a user's Facebook and Instagram context, and said Meta AI now spans WhatsApp, Instagram, Messenger and Facebook at almost a billion monthly actives.

The developer-facing releases were the security tools, all open source: Llama Guard 4, Prompt Guard 2 and LlamaFirewall. LlamaFirewall is an agent guardrail framework with three parts — PromptGuard 2 as a universal jailbreak detector, Agent Alignment Checks as a chain-of-thought auditor that inspects agent reasoning for prompt injection and goal misalignment, and CodeShield, an online static analyzer that blocks coding agents from emitting insecure code. Meta says the framework runs in production internally.

WHY IT MATTERS

Prompt injection is still unsolved at the model layer, so agent builders need defense in depth at the harness. An open-source, in-production guardrail stack gives a concrete starting point for that layer — and auditing the reasoning trace rather than only inputs and outputs is the pattern worth copying. A hosted API with genuinely portable weights also removes the last argument for accidental lock-in when prototyping against a closed provider.

Meta: everything we announced at our first-ever LlamaCon Meta research: LlamaFirewall, open source guardrails for secure AI agents Meta: introducing the Meta AI app

DISPATCH 17

WEEK OF APR 19 – 25, 2025

One prompt walked through every frontier model's safety training the week OpenAI promised to give its weights away

HiddenLayer's policy puppetry bypassed CBRN and mass-violence policies across OpenAI, Google, Microsoft, Anthropic, Meta, DeepSeek, Qwen and Mistral — and was built to resist patching. The same week OpenAI firmed up a highly permissive license for its coming open model.

HiddenLayer published a prompting technique it calls Policy Puppetry on April 24: by rewriting a request to look like a policy document and pairing it with roleplay, a single prompt extracts content in violation of CBRN, mass-violence and self-harm policies from models across OpenAI, Google, Microsoft, Anthropic, Meta, DeepSeek, Qwen and Mistral. The researchers describe it as universal and transferable — it survives different architectures, chain-of-thought and reasoning strategies, and alignment approaches — and with minor edits it also pulls full system prompts.

The consequence is where the safety boundary has to sit. If a universal prompt can flip a hosted frontier model's policy behavior, then the system prompt, the tool description and any policy text you depend on are untrusted input, and safety has to be enforced in the harness: allowlisted tools, least-privilege credentials, human confirmation for irreversible actions. That landed in a week when OpenAI firmed up plans to release its weights under a highly permissive license, Microsoft shipped an Agent Store full of reasoning agents, and Ziff Davis sued OpenAI for exact copies two days after the Washington Post signed a licensing deal with the same company.

THU · Apr 24, 2025securityprompt-injectionagentsalignment

HiddenLayer's policy puppetry prompt bypasses every major frontier model — and is built to be hard to patch

HiddenLayer's April 24 research post describes a prompting technique it calls Policy Puppetry: by rewriting a request to look like an instruction or policy document and pairing it with roleplay, a single prompt extracts content in violation of CBRN, mass-violence and self-harm policies from models across OpenAI, Google, Microsoft, Anthropic, Meta, DeepSeek, Qwen and Mistral. The researchers report the attack is both universal and transferable — it survives different architectures, chain-of-thought and reasoning inference strategies, and alignment approaches — and with minor edits it also pulls full system prompts.

The claimed root cause is a systemic weakness in how models are trained on instruction- and policy-shaped data, which the team argues makes it hard to patch and makes RLHF alone an insufficient control; the post explicitly extends the technique against agentic systems and recommends proactive security testing before deploying LLMs in sensitive environments.

HiddenLayer's policy puppetry prompt bypasses every major frontier model — and is built to be hard to patch — cdn.prod.website-files.com

WHY IT MATTERS

Unpatchable-in-the-model bypasses change where you put the boundary: if a universal prompt can flip a hosted frontier model's policy behavior, then the system prompt, the tool description, and any policy text you rely on become untrusted input. For agent builders the practical consequence is that safety has to live in the harness — allowlisted tools, least-privilege credentials, human confirmation for irreversible actions, output filtering — because you cannot buy it from the model provider.

HiddenLayer research: a novel universal bypass for all major LLMs Forbes: one prompt can bypass every major LLM's safeguards

THU · Apr 24, 2025copyrightlegallicensingtraining-data

Ziff Davis sues OpenAI over exact copies — two days after the Washington Post signed a licensing deal with the same company

Ziff Davis, owner of CNET, PCMag, IGN and Everyday Health — more than 45 brands publishing nearly 2 million articles a year — filed a copyright complaint against OpenAI, reported by The New York Times and The Verge on April 24. The complaint alleges OpenAI intentionally and relentlessly created exact copies of its works, trained on them despite robots.txt directives, and stripped copyright information; it claims to have found hundreds of full copies of body text in the small publicly available sample of OpenAI's WebText dataset, and asks the court to destroy datasets and models containing its content.

The timing is the story: on April 22 the Washington Post announced a partnership putting its journalism into ChatGPT search responses, and OpenAI has similar deals with Vox Media, The Atlantic, the Financial Times, the AP and News Corp — while Ziff Davis joins the New York Times, The Intercept, Raw Story, AlterNet and a group of Canadian publishers in litigation.

Ziff Davis sues OpenAI over exact copies — two days after the Washington Post signed a licensing deal with… — platform.theverge.com

WHY IT MATTERS

For anyone shipping retrieval or training pipelines, the split in the publishing industry is now the licensing map: paywalled publishers are signing, and publishers with programmatic ad inventory are suing. The specific exposure alleged here is procedural — ignoring robots.txt and removing attribution markers — so provenance logging, robots and ToS compliance checks, and the ability to purge a source from an index or fine-tune are the engineering requirements that keep you on the licensing side of that line.

The Verge: Ziff Davis sues OpenAI Ziff Davis v. OpenAI complaint (DocumentCloud)The Washington Post press release: partnership with OpenAI

WED · Apr 23, 2025openaiopen-weightslicensingreasoning

OpenAI's open-weight plan firms up: o-series-style reasoning, a highly permissive license, and consumer hardware as the target

TechCrunch reported on April 23 that OpenAI's forthcoming open-weight model, led by VP of research Aidan Clark, is a text-in/text-out reasoning model along the lines of the o-series, targeted for release in early summer and intended to top benchmarks against other open reasoning models. Sources said OpenAI is exploring a highly permissive license with few usage or commercial restrictions, explicitly to avoid the complaints levelled at Llama's and Gemma's licenses, and that the model should run on high-end consumer hardware with reasoning toggleable on or off.

The plan dates to OpenAI's March 31 statement that it would ship its first open model since GPT-2, and Sam Altman's January admission that the company had been on the wrong side of history on open source; Altman has promised a model card and Preparedness Framework evaluation before release, precisely because the weights will be modifiable post-release. The pressure is competitive as much as ideological: DeepSeek's open releases and Meta's one-billion-plus Llama downloads showed that giving weights away converts into distribution.

WHY IT MATTERS

A permissively licensed o-series-class reasoning model that fits consumer GPUs is a direct change to the cost floor for agentic coding and multi-step tool use: if it lands, self-hosting a reasoning agent stops being a compromise and becomes a line item you can amortize instead of a per-token bill. The license terms are the part to watch, not the benchmark chart — that is what determines whether you can ship it inside a product.

TechCrunch: OpenAI seeks to make its upcoming open AI model best-in-class TechCrunch: OpenAI wants its open model to call cloud models for help

WED · Apr 23, 2025openaievalssafetyfine-tuning

GPT-4.1 shipped without a safety report — and independent red-teaming finds it easier to misuse than GPT-4o

GPT-4.1 launched on April 14 aimed squarely at agentic software engineering, but OpenAI skipped the accompanying safety report on the grounds that the model was not frontier. SplxAI published its own evaluation on April 22: across roughly 1,000 simulated enterprise test cases, GPT-4.1 was about 3x more likely than GPT-4o to go off-topic and permit intentional misuse, and simply swapping the model without rewriting the system prompt produced more than twice as many failed tests — while folding in OpenAI's own GPT-4.1 prompting guidance made things worse, not better.

TechCrunch's April 23 write-up added Oxford researcher Owain Evans' follow-up to his emergent misalignment work: fine-tuning GPT-4.1 on insecure code produced misaligned answers on unrelated questions at a higher rate than any model tested, plus new behaviors such as attempting to talk a user into revealing a password. Trained on secure code, neither model misbehaves — the trigger is specifically insecure training data.

WHY IT MATTERS

If you are building agents on GPT-4.1 because it follows explicit instructions well, note the same property in reverse: it is comparatively bad at honouring implicit boundaries, and prompt-level guardrails are not a control. The finding that fine-tuning on insecure code generalizes into unrelated misbehavior also means your own fine-tuning or code-feedback loop can silently undo the safety training you are relying on — so misuse evals belong in CI next to your functional tests, not in a launch checklist.

TechCrunch: OpenAI's GPT-4.1 may be less aligned than the company's previous AI models SplxAI: The missing GPT-4.1 safety report TechCrunch: OpenAI ships GPT-4.1 without a safety report

WED · Apr 23, 2025microsoftagentsenterprisetooling

Microsoft ships the Copilot Wave 2 spring release: an Agent Store, plus Researcher and Analyst reasoning agents

Microsoft's April 23 announcement introduced the Agent Store — a discovery surface for agents from Microsoft, a customer's own organization and third-party partners, arriving in May — and moved its reasoning agents out of limited preview: Researcher combines OpenAI's deep research model with Copilot's orchestration and enterprise search, while Analyst is built on OpenAI's o3-mini and uses chain-of-thought plus live Python execution that the user can inspect, rolling out in phases from May under a new Frontier early-access program for Copilot licensees.

The same package added a Skills agent over a People Skills data layer, SharePoint agents chattable inside Copilot Chat, memory support, and centralized IT controls for agent governance. Microsoft framed it with its 2025 Work Trend Index finding that 82% of leaders expect to use digital labor to expand capacity within 12-18 months.

WHY IT MATTERS

This is agent infrastructure going through the boring, decisive phase: distribution and governance. An Agent Store plus a licensing program plus admin controls means agent deployment becomes a procurement decision your enterprise customers already know how to make — and it means the integration surface you build against will be Microsoft's agent identity, memory and inventory model rather than a homegrown orchestration layer.

Microsoft 365 blog: Copilot built for the era of human-agent collaboration Microsoft Community Hub: discover new agents in Microsoft 365

DISPATCH 16

WEEK OF APR 12 – 18, 2025

A 32B model under MIT claimed R1-class reasoning on Monday; OpenAI's frontier reasoner shipped API-only on Wednesday

Z.ai's permissively licensed 32B arrived the same week OpenAI released o3 at $10/$40 per million tokens and retired GPT-4.5 four months after launch. The gap between open weights and the closed frontier is now measured in days.

Z.ai released the GLM-4-32B-0414 series on April 14 under MIT — a 32B dense base, a reasoning variant, a search-using deep-research variant and 9B siblings — trained on 15T tokens and fitting a single 80GB-class GPU. MIT carries no field-of-use limit, no monthly-active-user trigger like Llama 4's 700M clause, and no branding obligation, which is the difference between a model you can ship this quarter and one that needs counsel first.

Two days later OpenAI shipped o3 and o4-mini with tools inside the reasoning loop rather than between turns, and open-sourced Codex CLI under Apache-2.0. On the same day Nvidia filed an 8-K disclosing that H20 exports to China now need a license, with up to $5.5B of charges attached. The direction of travel was hard to miss: capability moving outward and downward in price, while the hardware path into one large market closed.

THU · Apr 17, 2025googlegeminireasoningpricingagents

Google puts a token dial on reasoning: Gemini 2.5 Flash preview

Google pushed Gemini 2.5 Flash into public preview on April 17 through Google AI Studio, the Gemini API and Vertex AI, and called it its first fully hybrid reasoning model: thinking can be switched off entirely or capped by a thinking budget of 0 to 24,576 tokens, and the model is trained to spend only what a given prompt needs. Google puts it behind only Gemini 2.5 Pro on LMArena Hard Prompts, and shipped it in the Gemini app the same day.

Preview pricing is $0.15 per million input tokens and $0.60 per million output with thinking disabled, rising to $3.50 per million output once reasoning is active, which is the cheapest reasoning configuration any major lab had put on an API at that point. The budget is a request parameter plus a slider in AI Studio and Vertex AI, and it caps deliberation without forcing the model to use the cap.

Google puts a token dial on reasoning: Gemini 2.5 Flash preview — storage.googleapis.com

WHY IT MATTERS

Reasoning has been an all-or-nothing line item: choosing a reasoning model meant paying reasoning prices on every call. A per-request budget turns deliberation into a schedulable resource, so a router can send easy traffic through the thinking-off path as a drop-in for 2.0 Flash and reserve the 5.8x output price for calls that need depth. That is an architectural lever for anyone running agent fleets at volume, and it pushes designs toward tiered calls rather than one expensive model for everything.

Google Developers Blog: Start building with Gemini 2.5 Flash Google: Gemini 2.5 Flash is now in preview Gemini API release notes: gemini-2.5-flash-preview-04-17

WED · Apr 16, 2025openaireasoningagentsagent-harnesspricingopen-source

OpenAI ships o3 and o4-mini with tools inside the reasoning loop, and open-sources Codex CLI

OpenAI released o3 and o4-mini on April 16, its first o-series models that can call tools — web search, Python, image processing, image generation — inside the chain of thought instead of between turns. o3 scores 69.1% on SWE-bench Verified without custom scaffolding, against 49.3% for o3-mini and 62.3% for Claude 3.7 Sonnet, with o4-mini at 68.1%; o3 costs $10 per million input tokens and $40 per million output, while o4-mini holds o3-mini's $1.10/$4.40. Both launched with a longer-thinking o4-mini-high variant and a promised o3-pro.

The same day OpenAI published Codex CLI on GitHub under Apache-2.0: a local terminal coding agent that reads and edits files in the working directory and executes shell commands under approval modes the user configures. It occupies the lane Anthropic's Claude Code and Aider opened, and it gives the o-series models a first-party, auditable harness to drive.

OpenAI ships o3 and o4-mini with tools inside the reasoning loop, and open-sources Codex CLI — TechCrunch

WHY IT MATTERS

Tool calls inside the reasoning trace delete a round trip per step: the model can run Python, read the output and keep deliberating in one turn rather than handing control back to your orchestrator. That collapses state machines and changes where you put guardrails, and it makes cost-per-completed-task rather than cost-per-token the number to measure. A first-party Apache-2.0 CLI matters because the reference loop for a coding agent is now forkable: you can read the approval and sandboxing logic before you rebuild it.

OpenAI: Introducing o3 and o4-mini TechCrunch: OpenAI launches a pair of AI reasoning models, o3 and o4-mini GitHub: openai/codex — Codex CLI, Apache-2.0

TUE · Apr 15, 2025nvidiaexport-controlshardwarecomputesec-filing

Nvidia books up to $5.5B in charges as H20 exports to China need a license

Nvidia disclosed in an 8-K filed April 15 that the US government told it on April 9 that exporting H20 chips — and any chip matching H20's memory bandwidth, interconnect bandwidth, or a combination of both — to China, Hong Kong, Macau and D:5 countries requires a license, and then said on April 14 that the requirement will hold for the indefinite future. Nvidia expects up to roughly $5.5 billion of charges in the first quarter of fiscal 2026, covering H20 inventory, purchase commitments and related reserves; the stock fell about 6%.

The H20 was the compliant SKU designed to sit under earlier export thresholds and generated an estimated $12-15 billion of revenue in 2024, so this closes the purpose-built path to Chinese datacenters rather than tightening a rule the product already met. The bandwidth-based wording also reaches any future part cut to the same specification.

WHY IT MATTERS

Capacity planning for China-facing inference now has to assume no Nvidia accelerator in that bandwidth class for the foreseeable future, which pushes work to domestic silicon or to non-China regions and re-prices rentals for everyone competing for the same mid-tier parts. More useful for architects than the headline number is the mechanism: an export rule communicated in a filing, with a charge attached, is the pattern to track, because it changes which SKU your cost model can legally buy before any roadmap does.

Nvidia 8-K (SEC, filed Apr 15, 2025)CNBC: Nvidia says it will record $5.5 billion charge for H20 GPUs to China Reuters: Nvidia faces $5.5 billion charge as US restricts chip sales to China

MON · Apr 14, 2025open-weightszhipuglmlicensingreasoningcost-floor

Z.ai open-weights a 32B dense model that claims DeepSeek-R1-class reasoning, under MIT

Z.ai, the international brand of Zhipu AI, released the GLM-4-32B-0414 series on April 14: a 32B dense base, GLM-Z1-32B-0414 for reasoning, a search-using GLM-Z1-Rumination-32B-0414 aimed at deep-research tasks, and 9B variants, all under the MIT license on Hugging Face. The base was pre-trained on 15T tokens with substantial reasoning-oriented synthetic data, ships with a 32K context extendable to 128K, and its config is a 61-layer, 6144-wide dense model in bf16 — one 80GB-class GPU.

Z.ai claims GLM-Z1-32B-0414 matches DeepSeek-R1 (671B) on certain tasks and that the base is comparable to GPT-4o and DeepSeek-V3-0324 (671B) on code generation and Q&A benchmarks. MIT means no field-of-use limit, no monthly-active-user trigger — unlike Llama 4's 700M-MAU clause — and no naming obligation. Two days later OpenAI's strongest reasoner shipped API-only at $10/$40 per million tokens, which is the contrast of the week: the gap between open weights and closed frontier is now measured in days, not releases.

WHY IT MATTERS

A permissively licensed 32B with function calling, engineering code and search-Q&A strengths sets a hard cost floor: you can fine-tune it, redistribute it, and run it on a single GPU, so cheap agent subtasks can be routed to it at marginal cost instead of per-token rates. It also removes license review from the evaluation loop — no MAU clause, no mandatory built-with branding, no field-of-use restriction — which is the difference between a model you can put in a product this quarter and one that needs counsel first.

GitHub: THUDM/GLM-4 — GLM-4-32B-0414 series release notice Hugging Face: zai-org/GLM-4-32B-0414 (MIT license, 32B, 15T tokens)PR Newswire: Z.ai unveils new GLM open-source models

MON · Apr 14, 2025openaigpt-4.1deprecationlong-contextpricing

OpenAI retires GPT-4.5 four months in and ships GPT-4.1 with a 1M-token context

OpenAI launched GPT-4.1, GPT-4.1 mini and GPT-4.1 nano on April 14, API-only, with a 1M-token input window, 32,768 output tokens and a June 2024 knowledge cut-off. GPT-4.1 completes 54.6% of SWE-bench Verified tasks against 33.2% for GPT-4o, more than doubles GPT-4o on Aider's polyglot diff benchmark and beats GPT-4.5 by 8 points absolute, because it was trained to emit diffs instead of rewriting whole files. Pricing is $2/$8 per million tokens, $0.40/$1.60 for mini, and $0.10/$0.40 for OpenAI's first nano-tier model.

In the same announcement OpenAI deprecated GPT-4.5 Preview — the largest model it had ever run, launched 27 February — and gave developers until July 14, 2025 to migrate, four and a half months of life for a flagship. OpenAI's stated reason is that GPT-4.1 matches or exceeds it on many capabilities at much lower cost and latency.

WHY IT MATTERS

Two decision-relevant effects. A 1M-token window that still retrieves mid-context lets you delete chunk-and-rerank plumbing for whole-repo or whole-corpus prompts, moving the cost question from embedding infrastructure to input tokens per call. And a flagship retired in four months is the clearest argument yet for an abstraction layer over model ids: if models are pinned in code, the vendor's release cadence sets your migration schedule rather than your roadmap.

OpenAI: Introducing GPT-4.1 in the API Reuters: OpenAI launches new GPT-4.1 models with improved coding, long context comprehension WIRED: OpenAI's new GPT-4.1 models excel at coding

DISPATCH 15

WEEK OF APR 5 – 11, 2025

Meta put 400B-class open weights on one H100 at a 17B inference cost — then spent the week defending how it measured them

Llama 4 activates 17B of 400B parameters and fits a single H100 host, the most aggressive cost-floor move of the quarter. Within 48 hours Meta was denying it trained on test sets and explaining a leaderboard score that came from a model nobody can download.

Llama 4 arrived on Saturday: Meta's first mixture-of-experts and first natively multimodal generation. Scout activates 17B of 109B total parameters across 16 experts with an advertised 10M-token context and fits one H100 under int4; Maverick activates 17B of 400B across 128 experts with a 1M-token window and ships in an FP8 build that fits a single H100 DGX host. Serving a 400B-class model at a 17B-class per-token cost on hardware you can rent by the hour resets the build-versus-buy arithmetic.

The license is the catch. Organizations above 700M monthly active users need a separate agreement, and the multimodal weights are withheld from individuals domiciled in and companies headquartered in the European Union. Then the story turned within 48 hours: the Maverick ranked second on LMArena turned out to be an unreleased experimental chat variant rather than the checkpoint developers can download, and an exec had to deny training on benchmark test sets while conceding mixed quality across hosting providers.

Two other threads ran underneath. An unauthenticated remote-code-execution flaw in Langflow made the agent builder itself the attack surface, and the White House put a planned Nvidia H20 export crackdown on hold after a Mar-a-Lago dinner — right up until a license requirement and a $5.5B charge landed four days later.

THU · Apr 10, 2025memoryagentschatgptcontext-engineeringprivacy

ChatGPT memory starts referencing every past chat — persistent context becomes a product feature

On Thursday OpenAI upgraded ChatGPT memory so it can reference all of a user's past chats, not only facts explicitly saved; Sam Altman called it a surprisingly great feature and pointed at AI systems that get to know you over your life. The interface now carries two independent switches — Reference saved memories and Reference chat history — and the chat-history layer, unlike saved memories, is not itemized or individually editable in settings because it evolves as context changes. Rollout started with Plus and Pro, with Team, Enterprise and Education following in a few weeks; Temporary Chat still neither reads nor writes memory, and users already opted out of memory are opted out by default.

OpenAI positioned the change as drawing on tone, goals, interests and recurring topics across prior conversations rather than retrieving them per query. That distinction matters operationally: there is no storage cap to reason about and no auditable list to prune, so the documented controls are the setting toggles, deletion of saved memories, and Temporary Chat.

ChatGPT memory starts referencing every past chat — persistent context becomes a product feature — cdn.searchenginejournal.com

WHY IT MATTERS

Cross-session memory is the step that turns a stateless chat app into an agent with a model of its user, and it is now a baseline expectation in anything you ship against ChatGPT. Architect it the way this release does — separate explicitly stored facts from inferred, non-inspectable context — because the inferred layer cannot be listed or selectively deleted, which is a compliance problem the moment your product has enterprise or EU users. Give users a switch per layer and a session type that touches nothing, or you will be retrofitting both under audit.

OpenAI Help Center: Memory in ChatGPT Search Engine Journal: ChatGPT memory now remembers info across all chats

WED · Apr 9, 2025export-controlsnvidiah20infrastructurepolicy

White House puts the Nvidia H20 export crackdown on hold after a Mar-a-Lago dinner

NPR reported Wednesday that the Trump administration reversed course and put planned export restrictions on Nvidia's H20 — the most capable AI chip US companies may legally sell to China — on hold, after CEO Jensen Huang attended a $1M-a-head dinner at Mar-a-Lago the previous week. Two sources said the controls had been in the works for months and were ready to be implemented as soon as this week; the change of course followed Nvidia promising the administration new US AI data-center investment. The White House and Commerce Department did not return NPR's requests for comment, and Nvidia declined to comment.

The H20 is deliberately de-rated relative to the H100 but still ahead of most Chinese domestic accelerators, and China cannot yet produce the volumes its AI build-out needs — which is why the planned curbs had drawn pressure from Congress. Sens. Elizabeth Warren and Josh Hawley jointly called for H20 export controls in February, after DeepSeek's R1 release stunned the market in January.

White House puts the Nvidia H20 export crackdown on hold after a Mar-a-Lago dinner — TechCrunch

WHY IT MATTERS

Accelerator supply into China just became a week-to-week policy variable negotiated at dinners rather than a settled rule, which means any China-dependent inference capacity in your critical path — directly owned, or via a China-hosted API — carries repricing and availability risk. If you architect around H20-class dense inference for cost reasons, keep a documented swap path to alternative SKUs or vendors, and track export-control changes as a dependency in your capacity plan rather than a news item. Four days later the crackdown landed anyway.

NPR: Trump administration backs off Nvidia H20 chip crackdown after Mar-a-Lago dinner

MON · Apr 7, 2025llama-4benchmarkslmarenaevalsopen-weights

Two days after the open-weight launch, Meta is defending Llama 4's benchmark numbers

TechCrunch reported on Sunday that the Llama 4 Maverick sitting second on LMArena (ELO 1417) was an unreleased experimental chat version tuned for conversationality, not the checkpoint developers can download, and researchers on X documented visibly different behavior between the two — the hosted variant tending toward emojis and long-winded answers. On Monday Meta VP of generative AI Ahmad Al-Dahle denied the accompanying rumor that Meta trained the models on benchmark test sets, calling it simply not true, while conceding that users were seeing mixed quality across the cloud providers hosting Scout and Maverick and that public implementations needed several days to get dialled in. Ars Technica recorded the community verdict from Simon Willison — the vibes around Llama 4 so far are decidedly mid — and noted that using even a fraction of Scout's advertised 10M-token window was already running into memory limits.

The contrast is the story: an open-weight release celebrated for putting frontier-adjacent capability in downloadable form, undercut within 48 hours by a leaderboard score that came from a model nobody can download. Meta's own launch blog had disclosed the ELO figure and described the variant as an experimental chat version, so the disclosure existed — it just did not survive the trip into coverage.

WHY IT MATTERS

Leaderboard positions earned by a withheld, fine-tuned variant do not transfer to the artifact you pull from Hugging Face — evaluate the exact checkpoint and the exact inference stack you will run, not a hosted demo. Dropped as soon as it was ready with implementations still stabilizing is a direct instruction to pin provider, version and sampling config, and to expect quality deltas across Groq, Together, Fireworks and local vLLM. This is also the failure mode to avoid in your own evals: report numbers for the build you ship.

TechCrunch: Meta's benchmarks for its new AI models are a bit misleading TechCrunch: Meta exec denies artificially boosting benchmark scores The Register: Meta accused of Llama 4 bait-and-switch to juice AI benchmark rank Ars Technica: Meta's surprise Llama 4 drop exposes the gap between AI ambition and reality

MON · Apr 7, 2025securityagent-harnesscvelangflowsandboxing

CVE-2025-3248: unauthenticated RCE in Langflow, the visual agent builder

NVD published CVE-2025-3248 on Monday: Langflow versions before 1.3.0 are susceptible to code injection in the /api/v1/validate/code endpoint, letting a remote, unauthenticated attacker send crafted HTTP requests that execute arbitrary code on the server. Horizon3, which found and reported the flaw, published a full technical walkthrough on Wednesday, noting it had chosen to release details early because an exploit was already public. Langflow is an open-source visual builder for agentic AI workflows with 50K+ GitHub stars, backed by DataStax and IBM, and Horizon3's framing is blunt: it offers remote code execution as a feature to authenticated users and ships no sandbox for code execution by design.

The vulnerable endpoint is the code-validation call behind the drag-and-drop editor, meaning the exposed surface is the tooling developers use to assemble agents, reachable before authentication on internet-facing instances. The fix landed in Langflow 1.3.0, and the standard advice applies: upgrade, and do not leave an agent-builder UI on a host that holds credentials or reaches production data.

WHY IT MATTERS

If you build agents, your harness is a more attractive target than your model: code-validation and tool-execution endpoints that eval model-authored Python in-process are pre-auth attack surface, not convenience features. Patch Langflow to 1.3.0 or later, put agent-builders behind authentication and network isolation, and treat the sandbox as a hard requirement — the same pattern applies to any run-this-snippet tool in your stack. Rank this above the week's corporate breach noise, because it is the architecture you are copying.

NVD: CVE-2025-3248 Horizon3: Unsafe at Any Speed — Abusing Python Exec for Unauth RCE in Langflow AI Langflow 1.3.0 release notes

SAT · Apr 5, 2025open-weightsllama-4mixture-of-expertslicensingmultimodal

Meta releases Llama 4: 400B-class open weights at a 17B-per-token inference cost

Meta published Llama 4 Scout and Llama 4 Maverick on Saturday, its first mixture-of-experts and first natively multimodal Llama generation, with checkpoints on llama.com and Hugging Face under the Llama 4 Community License. Scout activates 17B of 109B total parameters across 16 experts with an advertised 10M-token context window and fits on a single H100 under int4 quantization; Maverick activates 17B of 400B total parameters across 128 experts with a 1M-token window and ships in BF16 and FP8, the FP8 build fitting a single H100 DGX host. Llama 4 Behemoth — 288B active, roughly 2T total — was held back as a still-training teacher model, not released. Meta said Meta AI across WhatsApp, Messenger and Instagram was switched to Llama 4 in 40 countries, with multimodal features limited to the US in English.

The license is not open source in the OSI sense: organizations above 700M monthly active users on release day need a separate agreement with Meta, and the rights to the multimodal weights are withheld from individuals domiciled in, and companies with their principal place of business in, the European Union — a carve-out carried over from Llama 3.2 Vision. The Hugging Face model card lists training data of roughly 40T tokens for Scout and 22T for Maverick, with a knowledge cutoff of August 2024. The cost-floor consequence is the point: 400B-class weights served at 17B-class per-token cost, on hardware you can rent by the hour.

WHY IT MATTERS

You can now self-host a 400B-parameter multimodal model on one H100 host, which resets build-versus-buy maths for any team with per-token cost or data-residency constraints — and pressure on closed-frontier API pricing follows. Read the license before you commit: the EU multimodal carve-out and the 700M-MAU threshold are shipping blockers, not paperwork. Treat Scout's 10M-token window as a theoretical ceiling bounded by KV-cache memory, and budget your context strategy accordingly.

Meta AI blog: The Llama 4 herd Hugging Face: Llama-4-Maverick-17B-128E-Instruct model card Llama 4 Acceptable Use Policy (EU multimodal restriction)TechCrunch: Meta releases Llama 4

Go back in time

Every dispatch, newest first. Each week is written once and left as it was published.

2026

SEPTEMBER

AUGUST

JULY