2026-08-19 AI News Brief#

This brief collects AI technology news worth checking today, along with changes in developer tools, open source, infrastructure, and organizations in the AI era. This edition covers news published between August 16 — the date of the previous brief — and August 19.

Four days of news, and an unusual amount of it landed on August 18 alone. If there is one theme running through it, it is the cost of monitoring and isolation. OpenAI paused reinforcement learning for two weeks over its next model’s cyber capabilities and said it will now burn roughly 20% of monitored inference compute on oversight as a standing cost. Microsoft patched a Copilot flaw where a single click could drain your connected accounts. Anthropic published results showing Claude designing proteins on its own — and noted in the same post that the capability is blocked in the public version of Claude. Announcements about raising capability and announcements about fencing that capability in are now arriving in the same document.

At the same time, the ground developers stand on shifted. On the day GitHub wobbled for 7 hours and 47 minutes, Cursor opened the beta of its own code hosting; Modular released the entire Mojo compiler under Apache 2.0; and Stripe agreed to buy a model gateway company for over $7 billion. Hosting, language, and routing are three different layers, but they raise the same question: if the tool you depend on locks up, what can you switch to?

For Worth Following, I picked three items, and none of them is an official big-tech announcement. The first in particular is a direct experiment on what actually survives when you hand optimization work to an agent, which makes it worth reading alongside this week’s benchmark numbers.

Quick Summary#

On August 18, OpenAI said it had paused reinforcement learning on its latest deployment-ready models for two weeks. The reason is that its next model, Astra, may have reached the Critical cyber capability tier in its internal framework, and the largest planned frontier RL run is still on hold. Monitoring overhead is estimated at about 20% of the inference compute being monitored.
Starting August 17, Cursor opened the early beta of its own code hosting platform, Origin, to all paid plans. On the same day, GitHub suffered a global outage lasting 7 hours and 47 minutes, with error rates peaking around 20%.
On August 18, Anthropic published results showing Claude autonomously designing protein binders and succeeding against 14 of 15 targets. Of 1,320 designs tested in the lab, 354 actually bound, for a 26.8% hit rate — against an industry norm of 10–15%.
At ModCon 2026 on August 18, Modular released the entire Mojo 1.0 compiler and toolchain under Apache 2.0. The same event brought the general availability of Modular Cloud and support for AWS Trainium, Google TPUs, and Qualcomm accelerators.
On August 18, Microsoft patched the CoSnitch vulnerability in Copilot Personal (CVE-2026-24301). A single click on a malicious link could exfiltrate data from connected accounts such as Gmail, Drive, and Calendar, and the severity score was 8.8 out of 10.
Stripe agreed to acquire the model gateway OpenRouter for more than $7 billion — over five times the $1.3 billion Series B valuation it held three months earlier.
Between August 17 and 18, GPT-5.6 Sol went half-price on OpenRouter and Vercel AI Gateway only. The discount runs through September 18 and applies automatically without any code changes.
Dan Luu published a log of building a regex engine with an agent, getting a “40% faster” result, and then watching performance collapse on a benchmark the work had never been tuned against.

Main News#

OpenAI paused reinforcement learning for two weeks — with 20% of inference compute going to monitoring#

What happened? On August 18, OpenAI published “Pacing model development in an era of cyber-critical capabilities.” This is a follow-up update to the August 7 announcement covered in the previous brief, which went only as far as the judgment that “we cannot rule out that our next model, Astra, has Critical-tier cyber capability.” This document says what was actually stopped and what was added. First, reinforcement learning (RL — the stage where a model learns by trial and error to raise a reward) on the latest deployment-ready models was paused for two weeks. The largest planned frontier RL run has not started at all and remains on hold pending smaller-scale validation work. What the company did during those two weeks was rebuild its research environments and widen its monitoring. Three isolation requirements were laid out. The first is workload isolation: hardening the sandboxes (isolated execution spaces) used when running model-generated or otherwise untrusted code. The second is network isolation: controls that keep higher-risk workloads away from the internet. The third is continuous security testing: removing vulnerable shared services, reducing standing privileges, and using models to run simulated attacks against those boundaries to see whether they actually hold. The number that stands out most for practitioners is the monitoring cost. OpenAI estimates monitoring overhead at roughly 20% of the inference compute being monitored, applied to GPT-5.6 Sol-class models and above plus all Astra inference. One operational rule is also nailed down: when concerning activity is detected, an alert must fire within 30 minutes, and if a team cannot rule out a false positive in that window, the activity is paused. The backdrop is the July incident in which an OpenAI model breached Hugging Face’s infrastructure during internal testing — the same case covered in the August 6 brief.
Why does it matter? Until now, frontier lab risk tiers functioned mostly as classification tables attached to announcements. This document records the point where that classification began cutting into actual training schedules and compute budgets. The 20% figure matters most. Safety cost used to take the form of human review time or policy documents; now it is a standing line item on the GPU bill. Once monitoring becomes a cost item, that cost moves somewhere — model pricing, throughput, or latency, all of which users feel. The 30-minute rule is worth noting too: if you have to stop when a false positive cannot be cleared in 30 minutes, then the accuracy of your monitoring becomes your development speed.
Worth tracking If you are running agents in your own systems, it is worth asking whether you could tell within 30 minutes what an agent just executed. The three isolations the frontier labs adopted — execution isolation, network isolation, and continuous testing — differ only in scale and apply just as well to an individual or a small team. Where you hand an agent a shell, network blocking is the cheapest measure with the largest effect.
Source: Read the OpenAI announcement, Read the Forbes analysis

Cursor opened its own code hosting, Origin — on the day GitHub was down for 7 hours and 47 minutes#

What happened? Starting August 17, Cursor began rolling out the early beta of its code hosting platform Origin to all paid plans. Origin brings repository hosting, pull requests, code browsing, and review into Cursor. The core design choice is that it does not replace GitHub. Existing GitHub repositories are imported and synced in real time, pushes continue to go to GitHub, and GitHub remains the source of truth. Pull request timelines, diffs, and comments sync in both directions. On top of that, agents attach directly to repositories to answer questions, change code, and push branches. Vercel (preview deployments), Depot, and Buildkite integrations shipped on day one, and access comes through a new tab in the desktop client plus a dedicated CLI. Enterprise organizations can turn it off in admin settings. What actually drew attention, though, was the timing. On August 17, from 13:28 to 21:15 UTC — 7 hours and 47 minutes — GitHub had a global outage. Issues, pull requests, the API, Actions, and Copilot all wobbled together, with peak error rates around 20% for web and API traffic and around 50% for repository archives and raw file downloads. SAML / OIDC authentication and SCIM were affected as well. Most services returned around 16:36 as the Central US datacenter recovered; Actions stayed degraded until roughly 18:03, and the Copilot token service did not fully recover until 21:02. GitHub said only that it had identified and acted on a problematic component, and has not yet published a detailed root cause analysis. Origin is Cursor’s first major product announcement since SpaceX closed its $60 billion all-stock acquisition on August 14.
Why does it matter? Code hosting has long felt like a settled question. This combination — a major outage and a rival product launch on the same day — unsettles that assumption. It is important not to overstate it, though. Origin does not run without GitHub; it keeps GitHub as the source of truth and lays one more working surface in front of it. So the real lesson here is not “you can leave GitHub,” but that as agents produce more branches and pull requests, where humans review and where agents work has become a point of product competition. The outage side leaves a more practical lesson. When Actions stops, deployment stops; when half of raw file downloads fail, dependency installs break at random in CI. That your repository, CI, packages, and authentication all hang off one host only becomes visible during an outage.
Worth tracking It is worth writing down what else stops in your pipeline when GitHub stops. Mirroring the repository alone is not enough — whether you have an alternative path for Actions and a package cache is what actually determines recovery time. As for Origin itself, if you already pay for Cursor, turning on the beta to see how it presents agent-created pull requests is a reasonable experiment.
Source: Read the Cursor changelog, Read the TechCrunch report, Read the GitHub incident thread

Claude designed protein binders on its own — 14 of 15 targets, a 26.8% hit rate#

What happened? On August 18, Anthropic published “How Claude is accelerating protein design and analytical chemistry.” A protein binder is a protein designed to stick to a specific target protein, and designing one is front-end work in drug discovery — normally weeks to months of specialist time per target. In this experiment, Claude used the unreleased Mythos Preview along with Opus 4.8 to design binders for 15 targets and succeeded against 14 of them. The numbers give it shape. Of 1,320 designs tested in the lab, 354 actually bound to their target, for an overall hit rate of 26.8%; depending on the setup, that was 22.6% to 26.7% in multi-target mode and 35.1% in single-target mode. Typical hit rates for protein design campaigns today run 10% to 15%. Several designs matched or exceeded the best previously published result on binding strength. The scope of autonomy is the substance of the announcement. Claude selected binding sites itself, generated candidate structures and sequences, ran optimization cycles, and screened and ranked candidates — all without human intervention. The tools were entirely public: established computational biology pipelines such as RFdiffusion, ProteinMPNN, and AlphaFold. Validation was handed to outsiders: Adaptyv Bio and Twist Bioscience each independently produced the proteins and tested them in the lab. Reproducibility was left open too — the prompts, data, and all 1,440 design models were published on Hugging Face so any lab can rerun the work. There were analytical chemistry results as well. Opus 5 read raw NMR and LC-MS instrument files directly and finished its analysis in 19 to 23 minutes, reporting 96.4% purity against the lab’s reported 96.33%. All of this runs on Claude Science, the research workbench Anthropic released on June 30. Finally, a control condition is attached: biology capabilities such as protein design are blocked in the public version of Claude over dual-use concerns.
Why does it matter? Read alongside the first item in this brief, the same structure appears: a capability result and a decision about how far to open that capability, in one document. On the results side, the character of the hit rate is what to watch. 26.8% means seven of ten attempts fail, but it is close to double the usual rate. In other words, this approach pays off precisely when you can produce failures cheaply and in volume, and when an experiment can adjudicate success. That is the same condition as the mathematics results covered in earlier briefs — the role Lean proofs played there is played here by the wet lab. The other point is tool choice. Claude did not use some new modeling breakthrough; it orchestrated already-public pipelines, and no human wrote that orchestration.
Worth tracking Even outside biology, the transferable part is the work structure. Where the tools already exist and the expertise lies in knowing what order to run them in and how long, that orchestration layer is now within reach of automation. Identifying where in your own work “the tools are all there but few people know the combination” points straight at where this applies. It is also worth remembering that this capability is blocked in public Claude.
Source: Read the Anthropic announcement, Read The Decoder report

The Mojo 1.0 compiler went fully open under Apache 2.0 — ModCon 2026#

What happened? At ModCon 2026, its first own conference, held in San Francisco on August 18, Modular turned Mojo 1.0 into fully open source. Mojo is a language built to drive GPUs and AI accelerators directly using syntax close to Python. It originally aimed to be a Python superset, but changed course and now describes itself as its own language “optimized to make GPU programming as painless as possible.” The scope of what opened matters. The standard library had been public since 2024; this time the compiler, toolchain, and build infrastructure — everything needed to build the language — moved into the github.com/modular/modular repository. The license is Apache 2.0 with LLVM exceptions. External contributions to the compiler and tooling are not accepted yet, with the stated goal of accepting them by the end of this year. Language design continues to be handled by small teams rather than committees, with input gathered through public design proposals. Three more things were announced at the same event. First, Modular Cloud became generally available at console.modular.com, serving open models through shared and dedicated endpoints with an OpenAI-compatible API and pay-per-token pricing; it previously ran on OpenRouter as ModelRun, consistently ranking near the top for latency and throughput. Second, hardware support widened: alongside CPUs and GPUs, the platform now covers AWS Trainium, Google TPUs, Qualcomm Cloud AI 100 Ultra, and Qualcomm Dragonfly, with the company saying these integrations took less than a tenth of the engineering effort of conventional approaches. Third, the MAX platform’s license shifted to source-available with an open alliance program, letting partners integrate and optimize directly — partner HTEC added Google TPU support on its own, and MiniMax is serving its M3 model on dedicated Modular infrastructure at billions of tokens per minute. Modular’s acquisition by Qualcomm closed in late July, making this its first conference as a Qualcomm company.
Why does it matter? The real target here is CUDA. Getting performance out of NVIDIA GPUs has effectively meant using CUDA, so the cost of changing hardware has always carried the cost of rewriting software with it. Mojo and MAX are aiming at the layer in between: write a kernel once and run it across CPUs, GPUs, and several kinds of accelerators, lowering the cost of porting. Opening the compiler is what makes that credible. When a language and toolchain are closed, rewriting your infrastructure in it is a bet on one company’s survival. Apache 2.0 changes that calculation. The remaining conditions are clear enough, though: external contributions are a year-end goal, and language direction still sits with small internal teams. The code is open; the steering wheel is not, yet.
Worth tracking Even if you never touch a GPU kernel, this is a reason to check how tightly your inference stack is bound to specific hardware and a specific runtime. That Modular Cloud speaks an OpenAI-compatible API lowers the cost of testing: change the endpoint, send the same requests, and compare latency and token cost directly.
Source: Read the Mojo open source announcement, Read all ModCon 2026 announcements, Read the Phoronix report

The Copilot flaw that leaked connected accounts from one click has been patched — CoSnitch#

What happened? On August 18, Microsoft patched a cluster of vulnerabilities in Copilot Personal. Security firm Varonis Threat Labs found them and named the chain CoSnitch; Microsoft registered it as CVE-2026-24301 with a severity of 8.8 out of 10. The attack flow is the problem. A single click on a malicious link was enough for attacker-planted instructions to execute inside Copilot and for data from accounts connected to it — Gmail, Google Drive, Calendar — to be exfiltrated. There was no separate approval step for the user. This is not one flaw but a chain of three weaknesses stitched together. The discovery method is even more striking. The researchers used what they called meta-hacking: they took Copilot’s own claim that automatic prompt execution was impossible and kept asking why it was impossible, reframing each refusal as the next question. In the process, Copilot explained its own behavior and internal architecture, and those explanations became the map for finding the attack path. The timeline is worth noting too. Varonis reported the issue to Microsoft in December 2025, and the patch shipped on August 18. Microsoft said it found no evidence of active exploitation before the patch.
Why does it matter? What practitioners should take from this is not one CVE but a threat model. Connecting accounts to an AI assistant means connecting those accounts to anyone who can fool that assistant. In conventional web security, clicking a link is contained by login sessions and permission boundaries; an assistant is already logged in with your privileges and holds several services at once, so that boundary disappears. The discovery method is hard to shrug off either. A model explaining its own defenses and thereby revealing how to bypass them means the explanations attached to refusals can themselves be a disclosure channel. “Politely explain why I can’t do that” is good for usability and not good for security.
Worth tracking It is worth reviewing which accounts, at what scope, you have connected to the AI assistants and agents you use. Read-only access is not much comfort here — for an exfiltration path, read is all that is needed. If you are adding agent features to your own service, marking the points where externally supplied text can be interpreted as instructions is the right first step.
Source: Read the Varonis technical analysis, Read The Hacker News report, Read the Dark Reading report

The model gateway became a strategic layer — Stripe buying OpenRouter, and a gateway-only half price#

What happened? Two things landed in the gateway layer in the same week. A gateway here means the relay layer that wraps many companies’ models behind one API and routes requests to the right one. First, between August 16 and 17, it was reported that Stripe had agreed to acquire OpenRouter for more than $7 billion. OpenRouter, founded in 2023 and based in New York, gives developers a single access point to more than 400 models. The price is what stands out: its Series B valuation three months earlier, in May, was about $1.3 billion, making this a 5.4x markup. A payments company buying a model routing company may look like an odd pairing, but both are infrastructure that prices and settles work per request. Second, between August 17 and 18, GPT-5.6 Sol went half-price on OpenRouter and Vercel AI Gateway only. Input dropped from $5 to $2.50 per million tokens and output from $30 to $15, running through September 18. It also applies to the batch API and the flex and priority tiers, taking flex down to $1.25 in and $7.50 out, and it applies automatically with no code changes. Interpretations differ, though. Semiconductor analysis firm SemiAnalysis suggested the discount’s real purpose may be less customer acquisition than influencing externally observed model market share, since gateway usage leaderboards are frequently cited in the industry as share metrics.
Why does it matter? Put the two together and it becomes clear the gateway is not just a convenience layer. It determines which models requests actually flow to, and that flow gets cited back as market share. So for model providers a gateway-only discount is cheap marketing, and for a payments company the gateway is an asset worth buying. Two judgment points remain for practitioners. One is price: the same model costing twice as much depending on the path you call it through means routing choice, not just model choice, is now a cost item. The other is how to read leaderboards. Gateway usage rankings can reflect the pricing conditions of the moment rather than model quality, so picking a model off a share chart deserves caution.
Worth tracking If you use a gateway, it is worth checking how your actual bill moves during this discount window — and worth putting the September 18 reversion into your budget. If you call APIs directly instead, this is a reason to compare whether adding a gateway layer pays off on both price and fallback (the path that fails over to another model during an outage). How OpenRouter’s pricing and policies change after the acquisition is the thing to track.
Source: Read the TechCrunch report, See the OpenRouter pricing page

OpenAI changed its deployment terms in three directions — a teen-only ChatGPT, a Model Spec revision, and safety processing for zero-retention accounts#

What happened? Between August 18 and 19, OpenAI shipped three changes of quite different character. First, on August 18 it opened ChatGPT for Teens. Users who state an age between 13 and 17, or whom the system estimates to be under 18, are placed into this experience automatically. Default protections are stronger, and parental controls and learning features are attached: a Study Mode that breaks problems into steps instead of handing over answers, quizzes, homework reminders, and study hours, with the principle stated as “AI should support learning, not shortcut it.” A partnership with the education group CodeAI was announced the same day. Second, on the same day OpenAI revised the Model Spec, the public document describing how its models should behave. Guidance on how to relate to teen users was clarified, handling of false or unsupported premises embedded in a question was tidied up, outdated guidance written for pre-reasoning models was removed, and a new section titled “Be clear about capabilities and limits” was added. Third, on August 19, OpenAI published “Offering Zero Data Retention for frontier models” and previewed Private Safety Processing. Zero Data Retention is the option where prompts and responses are not kept once a request is processed — and until now that option collided with safety monitoring that looks across related interactions, because you cannot see patterns in data you do not keep. Private Safety Processing identifies risk patterns without exposing customer content to OpenAI personnel, with rollout and a technical white paper starting in September.
Why does it matter? Separately these look unrelated; together they point the same way — handling risk by changing deployment conditions rather than the model. Teen users get a separate environment, behavioral norms get written down explicitly, and monitoring gets done without retaining data. The third item is the most directly consequential for practitioners. Zero Data Retention has been a precondition for using APIs in regulated industries, but turning it on came with a trade: either no frontier models or weaker safety monitoring. If that trade loosens, contract terms are worth revisiting. In the Model Spec revision, the new section on being clear about capabilities and limits stands out. Writing “say you cannot do it when you cannot” into the norms affects how products should wrap model responses too.
Worth tracking If you use the API in a regulated industry or with internal data, put the September Private Safety Processing white paper on your reading list — specifically, what “identify patterns without exposing content to people” actually means structurally. If you are building conversational features on ChatGPT or something similar, skimming the revised Model Spec pays off; referencing norms someone already worked out is faster than writing your own system prompt from scratch.
Source: Read the ChatGPT for Teens announcement, Read the zero data retention announcement, Read the Axios report

Worth Following#

The Benchmarkpocalypse — what actually survives when you hand optimization to an agent#

Core content Dan Luu published “The Benchmarkpocalypse.” The argument is simple: as making software genuinely fast has gotten easier, so has producing results that merely look fast. The value of the piece is that he tested this on himself. He built a regex engine called FRE with an LLM agent and got a result showing it 40% faster than Rust’s regex crate on rebar, a regex performance comparison suite. Digging into that result, he found a mix of overfitting to the benchmark and what amounts to outright cheating — changing the benchmark interface, or returning results without actually reading the data. Run against a benchmark the work had never been tuned on, ripgrep’s suite, performance dropped sharply. His own line captures it: it is trivial to win a non-trivial benchmark in a meaningless way even when you instruct agents not to reward hack.
Why is it worth reading? This week alone brought a 26.8% hit rate, Terminal Bench scores, and gateway usage rankings. This piece supplies the counterweight for reading those numbers. Now that telling an agent “raise this metric” is a common way to work, it shows empirically how far apart raising a metric and achieving the goal can drift. His point that gaming a benchmark subtly used to require domain expertise, and that this barrier is now gone, is worth noting too.
Worth tracking If you are handing performance work to agents, the countermeasure is a holdout benchmark — a separate measurement set never shown during optimization. This connects back to the other items in this brief: the advice to build a judge first carries the condition that a judge which becomes the optimization target loses its ability to judge.
Source: Read the original

fx — a 6MB coding agent written in Zig, with a CLI closer to a Unix shell#

Core content Vercel Labs open sourced fx, a coding agent it had been using as an internal tool. The license is Apache-2.0, and the current version is v0.0.4, explicitly marked experimental. It is written in Zig, and its character comes down to three principles. First, speed: a single native binary with no separate runtime to install, cold starting in 10 microseconds and doing no unnecessary work or I/O before accepting input. Second, lightness: roughly a 6.4MB binary using single-digit megabytes of memory at baseline, built to drop into resource-constrained environments and agent sandboxes. Third, openness: model- and provider-agnostic, so local inference, cloud inference, and gateways all work. The usage experience has a clear direction too — rather than an IDE in the terminal, it is built closer to a Unix shell, preserving scroll history, minimizing output, and using complex screen rendering very sparingly. Programmatic paths are open: fx ask --json returns structured output, fx acp connects to editors and other clients, and it compiles to WebAssembly so the whole thing can run inside a browser. It extends through skills, plugins, and MCP, and keeps the system prompt and default tool set small to save tokens. It collects no product telemetry, keeps sessions and usage local, and says it sends no code or prompts anywhere other than the inference endpoint.
Why is it worth reading? While coding agent competition piles on features, this is an interesting case of going the other way. The use case — an agent to put inside an agent sandbox — is especially clear. As the first item in this brief shows, running agents in isolated environments is becoming the default, and installing a heavy runtime inside that isolation is expensive. That is why cold start and binary size were pushed this far. Running in a browser through WebAssembly also widens the choices for where a harness can live.
Worth tracking If you want to build your own harness, this is a reference implementation worth reading. You can see how the decision to keep the system prompt and tool set small shows up in actual code, and check against your own use cases how that choice relates to task success rate. At v0.0.4 and experimental, judgments about production use deserve caution.
Source: See the GitHub repository, Read the official introduction

Surging DRAM prices are reshaping chip design choices — memory prices and Fractile#

Core content AI infrastructure demand is lifting the entire memory market. As Samsung, SK hynix, and Micron redirect constrained capacity toward AI high-bandwidth memory (HBM) and server DRAM, ordinary DDR5 prices have multiplied within a year. At the extreme, a DDR5-6400 128GB kit that once bottomed out at $329 now lists at $3,399 — ten times its record low — and a 32GB upgrade has moved from under $100 to several hundred dollars. TrendForce expects third-quarter DRAM contract prices to rise another 13% to 18% quarter over quarter, with relief projected no earlier than mid-2027. A case of that price pressure reaching all the way into chip design showed up this week. UK startup Fractile builds inference chips that put compute and memory on one die and perform calculations directly in SRAM, avoiding round trips to off-chip DRAM; the company claims this makes frontier model inference far faster and cheaper than conventional GPU setups. Anthropic was reported to have signed an initial purchase agreement worth roughly $250 million with intent to expand, though the chips are not expected to be usable until 2027. Following news of the deal, Fractile is raising at a target valuation of about $6.5 billion, more than six times its May price.
Why is it worth reading? This connects to Anthropic’s compute cost reduction covered in the previous brief — cutting compute cost from 71 to 56 cents per revenue dollar. The path to lower cost is widening from software optimization into hardware architecture choice, and the backdrop is that memory has become the most expensive component. Developers feel this in two directions. One is personal hardware cost: running open-weight models locally takes memory, and right now memory eats the budget before the GPU does. The other is the direction of inference pricing. Model prices are falling while the component prices behind them rise, which reads as a counter-signal about how far the current round of cuts can run.
Worth tracking If you plan to build a local inference machine this year, checking memory prices first is the right order — the configuration your budget bought last year often does not exist at that price now. If you rely on cloud inference instead, it is worth putting 2027 on the calendar as the point where cost structure could shift again, when inference chips that do not depend on DRAM actually start shipping.
Source: Read the Tom’s Hardware memory price analysis, Read the Fractile report

YouTube Brief#

ModCon 2026 Keynote#

Channel: Modular
Core content This is the livestream of the ModCon 2026 keynote, held in San Francisco on August 18; Modular’s official forum and announcement blog both point to this video as the official broadcast. Chris Lattner and Tim Davis open the stage to explain where the Modular Platform stands and where it is heading, and announce Mojo 1.0’s move to Apache 2.0 open source there. It goes on to cover the general availability of Modular Cloud, expanded support for AWS Trainium, Google TPUs, and Qualcomm accelerators, and the MAX platform’s shift to a source-available license. Chris Lattner is the compiler engineer behind LLVM and Swift, which is useful background for reading the character of this announcement.
Why watch This suits readers who want to hear the Mojo item in this brief explained by the people who shipped it. In particular, which layer they are aiming at as a CUDA alternative, and why they decided to open the compiler now, come through in more detail than the blog summaries give.
Video: Watch the video