Apple will restrict macOS Full Disk Access, naming AI agents as the reason
Apple said it will add controls requiring explicit user action before an app can be granted Full Disk Access on macOS, citing the growing risk from capable AI agents. The announcement followed a report that Meta’s Muse agent read a columnist’s private messages, and Meta spent the same day open-sourcing the firmware and SDKs for building your own Muse hardware. Federal prosecutors charged a California company owner with routing more than $300 million in Nvidia-equipped servers to China through Malaysia and Singapore, and Kevin Mandia’s Armadin raised $255.5 million for autonomous offensive security.
Security #
Apple will require “very explicit user action” to grant Full Disk Access, citing AI agents #
Apple / TechCrunch / Ars Technica
Apple’s developer news post says some developers are using Full Disk Access “in ways that could put users at risk, exposing everything on their systems… without users’ full knowledge and understanding,” and that it will introduce additional controls so the permission can only be granted through very explicit user action. Full Disk Access is the macOS permission that covers files, mail, Messages and browsing history; it was designed for backup software and has been adopted by agent apps that want to read a user’s local context. Apple’s stated reason is prospective: “As AI agents become increasingly capable and autonomous, the risks associated with this level of access will grow substantially.” Two incidents preceded the post — an Inc. columnist reporting that Meta’s Muse agent read his private messages, and a Wired report on a flaw in ChatGPT’s Mac app that could have exposed sensitive data. No macOS version, date, or technical mechanism was given, and Apple did not answer TechCrunch’s questions.
This is the first time a platform vendor has narrowed an OS permission and named autonomous agents as the cause, and the framing matters more than the mechanism, which does not yet exist publicly. Full Disk Access is a single binary grant over the most sensitive corpus on a personal machine, and it was already a poor fit for software that acts continuously without supervision — an agent with FDA has standing read access to every credential, token and private conversation on the device, and nothing in the permission model distinguishes “back up my files overnight” from “read my Messages and decide what to do.” What Apple has announced is friction at the grant, not scoping after it, which addresses the consent problem and leaves the authority problem untouched: a user who deliberately grants access still hands over everything. Treat the whole thing as an intention until a beta ships, since there is no version number attached to it.
Circuit Breaker Labs runs up to hundreds of thousands of simulated user conversations a day against safety-critical chatbots #
TechCrunch
The five-person startup builds adversarial “crash-test dummy” agents that simulate users across ages, backgrounds, languages and cultures, then runs tens of thousands to hundreds of thousands of simulated interactions per day against a customer’s deployment and scores the responses. Co-founder Shirali’s stated premise is that benchmark-style inputs do not resemble real users: “Models are really good at handling standard speech patterns, but nobody actually talks like that,” so the simulations deliberately carry slang, coded language and typos. Human domain experts help construct the profiles, and the scoring is designed to produce auditable, explainable results. The target market is AI coaching, journaling and mental-health applications. No funding figure or customer names were disclosed.
The interesting claim is about distribution shift rather than about safety: a model that passes a red-team suite written in clean prose has been tested on a population that does not exist, and coded language is exactly the register in which the harmful cases arrive. Running the simulation continuously rather than at release is the other departure from how safety evaluation is usually bought — it treats a deployed chatbot as something whose behaviour drifts with every model update underneath it, which is accurate. Discount the volume figure accordingly: hundreds of thousands of simulated conversations a day is a statement about generation throughput, not about coverage, and nothing here establishes that the simulated population resembles the real one any better than the prose it replaces.
Regulatory & Policy #
DOJ charges a California company owner with smuggling more than $300 million in Nvidia-equipped servers to China #
US Department of Justice / Quartz / DigiTimes
Greg Lui, 38, of San Gabriel, who owns Earthmade Computer Inc. of City of Industry, California, was arrested on a three-count indictment: conspiracy to violate the Export Control Reform Act, outbound smuggling, and conspiracy to commit money laundering. Prosecutors allege that between 2023 and 2024 Lui and unnamed co-conspirators bought export-controlled servers containing Nvidia GPUs, submitted false documentation to US manufacturers misrepresenting the end destination, shipped the hardware to Malaysia and Singapore, and re-exported it to China without Commerce Department licences. Conviction carries up to 20 years on each conspiracy count and 10 years on the smuggling count.
The transshipment route is the part worth noting, because it is the same Malaysia-and-Singapore path that drove the 2025 licensing changes for those two countries, and this indictment covers conduct from 2023 and 2024 — so it is an enforcement action against the generation of diversion that the current controls were written in response to, not evidence about whether they work now. The mechanism alleged is documentary rather than technical: no evasion of any hardware control, just a false end-user statement to the manufacturer, which is the weakest link in an export regime that depends on self-certification at the point of sale. The money-laundering count is the one that usually does the work in these cases, since proving a server’s final resting place is harder than tracing what was paid for it. Everything here is an allegation in an indictment; nothing has been proven.
Open Source #
Meta open-sources Muse gadget firmware and SDKs under Apache 2.0 #
Meta / TechCrunch / Hacker News
Meta published two SDKs that let arbitrary hardware act as a client for Muse, its personal AI agent: an ESP32 Device SDK for microcontroller boards, covering display output, audio in and out, and sensor integration, and a Linux Device SDK that turns a Raspberry Pi or similar machine into a gadget capable of running custom commands for home automation and system administration. Firmware and SDKs are on GitHub under Apache 2.0, provided as-is without warranty. Reference hardware includes the Waveshare ESP32-S3 AMOLED display, M5Stack StickS3, Seeed reTerminal e-ink display, Raspberry Pi 5 and the Home Assistant Voice Preview Edition. A separate USB-C device, Muse Home Link, bridges Muse to a home network and smart-home devices over HTTP APIs and community skills; Meta is giving 5,000 units free to subscribers, shipping in a few weeks, with a Discord for support.
Giving away the client firmware is a cheap way to buy an ecosystem, and the cost structure explains it: every gadget built this way is a device that calls Meta’s inference on Meta’s terms, so the SDK is a distribution channel rather than a product. The thing to actually weigh before building on it is the trust boundary, which is new and undefined. A Raspberry Pi gadget that runs “custom commands for system administration” on behalf of a cloud agent is a remote-execution endpoint on a home network, and the community-skills model puts third-party code in the path between the agent and the commands. Read this next to Apple’s announcement above, which exists partly because Muse read someone’s messages: on macOS the agent’s reach is about to be narrowed by the platform, and on an ESP32 there is no platform to narrow it.
antirez releases ds4, an MIT-licensed C inference engine that runs frontier open-weight models on a 128GB Mac #
DwarfStar / Hacker News
Salvatore Sanfilippo, the creator of Redis, has published DwarfStar 4 — a “narrow” C inference engine targeting high-memory Macs, CUDA and ROCm machines, aimed specifically at running frontier-class open-weight models locally. Its central technique is asymmetric 2-bit quantization that compresses the routed experts of a mixture-of-experts model while leaving the paths every token traverses at higher precision; it also persists the KV cache to SSD so a session resumes without recomputing its prefix. On the reference machine, an M5 Max with 128GB, it reports 790.2 tokens/second prefill and 39.4 t/s generation at short context, falling to 398.5 t/s and 27.6 t/s at 65K. Supported models include DeepSeek V4 and V4.1 Flash, GLM 5.x and Qwen 3.8 Flash Next, with text and vision. It ships a CLI chat, OpenAI- and Anthropic-compatible local APIs, and a native coding agent, under MIT, for Apple Silicon machines with 64GB or more.
Quantizing routed experts harder than shared parameters is the right asymmetry for MoE and an underused one: a routed expert is consulted by a small fraction of tokens, so its error contributes proportionally less, while the always-on path is where precision loss compounds. That is what makes 2-bit tolerable here when 2-bit across the board is usually not. KV-cache persistence to SSD is the quieter feature and probably the more useful one for agent work, where the same long prefix is re-sent on every turn and prefill is the dominant cost — 398.5 t/s at 65K means a cold 65K prefill is still roughly three minutes of waiting that the cache removes. These are the author’s own numbers on one machine with no reported quality evaluation, which is the gap: a 2-bit expert quantization that costs nothing in benchmarks is a strong claim, and nothing here measures it.
Ai2 open-sources AstaBrief, an 8B report generator that writes a cited report in 51 seconds #
Allen Institute for AI / Hugging Face
AstaBrief 8B is built on Qwen3-8B and turns a research question plus retrieved literature excerpts into a cited scientific report, generating the whole report in a single pass rather than section by section. Its Fast mode averages 51.1 seconds per report against 178.5 seconds for Thinking mode, about 3.5x faster. On SQABench-CS2 it is competitive with Claude-powered pipelines on rubric scores, answer precision, citation precision and citation recall, and was benchmarked against Claude 3.5 and 3.7 Sonnet, o3, o4-mini, GPT-4.1 and DeepSeek models; DR-Tulu ranked highest on human preference for overall quality, while AstaBrief matched or exceeded alternatives specifically on citation accuracy. Among 374 Asta users, 29.1% used it on multiple days and 23% switched to Fast mode exclusively.
Citation accuracy is the metric that should decide a tool like this, and it is the one where an 8B model matching hosted frontier pipelines is genuinely surprising — citation precision and recall are mostly a question of whether the model will restrict itself to the excerpts it was given, which is a discipline rather than a capability. The honest reading of the full result set is a split decision: DR-Tulu wins on human preference, AstaBrief wins on grounding, so what Ai2 has released is the model you want when a wrong citation is worse than dull prose. The single-pass design is where the speed comes from and also where the limit sits, since a report written in one pass cannot revise an early section after reading a later source. The adoption numbers are small and self-reported from Ai2’s own product — 374 users is a pilot, not evidence of fit.
Model Releases #
Black Forest Labs ships FLUX 3 Image with bounding-box composition and open weights #
Black Forest Labs / Hacker News
FLUX 3 Image is the image generation and editing member of Black Forest Labs’ FLUX 3 family, which also spans video, audio and actions. The headline control surface is spatial: rather than describing a scene in prose, you draw a bounding box for each element on a 0-1000 coordinate grid and describe what goes in it. It renders natively at 2K and 4K, supports pixel-level editing that preserves the untouched regions of an original, composes from up to 10 reference images, and exposes layout planning for an agent to drive. Weights are available on Hugging Face and GitHub under both a non-commercial licence and a separate commercial weights licence for companies generating at scale. No benchmark numbers or pricing are published on the model page.
Coordinate-grid composition is the feature that matters for anyone generating images programmatically, because it converts a prompt-engineering problem into an API: a layout is a list of boxes and strings, which is something a calling program can compute, validate and diff, and a prose description of a scene is none of those things. That is also why the agent-driven layout planning is listed as a capability rather than a demo — the box list is the natural interface between a planner and a renderer. Two caveats. The dual-licence structure means “open weights” here is a commercial negotiation for the users most likely to want it, the same arrangement FLUX has used before. And the model page carries no evaluation at all, so there is nothing to compare against the prior generation except the feature list.
Infrastructure #
NVIDIA cuts DGX Spark to $4,999 with a 64GB variant that runs 100B-parameter models #
NVIDIA
A 64GB unified-memory configuration of DGX Spark arrives on 23 October from Acer, ASUS, Dell, Gigabyte, HP and MSI, starting at $4,999, keeping the GB10 Grace Blackwell Superchip, DGX OS and software stack of the 128GB model. NVIDIA says it runs models up to 100 billion parameters on one unit; two units linked over the built-in ConnectX-7 with a QSFP cable pool their memory to 128GB and reach 200-billion-parameter models at up to 1.7x the performance of a single system. A new Sync Cluster Assistant detects connected units, validates configuration and sets up the ConnectX-7 network automatically. The box ships with the NVIDIA Agent Toolkit, CUDA-X, Nemotron models and support for Ollama, vLLM and PyTorch, and a Sync Model Launcher arriving at the end of the month will one-click-deploy models such as Qwen 3.8 27B.
The number to hold is 1.7x from two units, because it is the honest version of the clustering pitch — doubling the hardware buys 70% more throughput, so the cluster exists to fit a model that does not otherwise fit, not to go faster. Halving the memory and cutting the price is a straightforward segmentation move, and at $4,999 the comparison is no longer against a cloud instance but against ds4 above on a Mac that someone already owns, which is a harder competitor than NVIDIA’s slide decks assume. Note also what “runs 100B-parameter models” omits: a parameter count is a capacity claim, and unified memory on a GB10 has nothing like HBM bandwidth, so decode throughput at that size is the figure that would decide whether this is usable and it is not in the announcement.
A Helion-based vLLM linear backend beats CUTLASS and DeepGEMM on Hopper, with over 10% end-to-end throughput on some workloads #
PyTorch / Red Hat / Meta
Red Hat and Meta engineers integrated Helion — a PyTorch-native, hardware-agnostic kernel DSL — into vLLM’s linear backend, expressing Standard GEMM, Split-K and Swap-AB as tunable parameters of one kernel rather than three hand-written kernels plus dispatch heuristics, and letting an ahead-of-time autotuner pick the variant and config per input shape. On an H100 80GB across Qwen3 1.7B through 32B and Qwen3.8-27B, geometric-mean kernel speedups are 1.110x for FP8_Dynamic and 1.178x for W8A8_INT8 over CUTLASS, and 1.149x and 1.177x for Block_FP8 over FlashInfer and DeepGEMM. End-to-end serving gains are consistent, exceeding 10% throughput on some workloads. The design only dispatches to Helion below 32 tokens and under CUDA Graph replay, falling back to the default kernel above that, which both avoids Helion’s per-launch CPU overhead and limits how many shapes need tuning. Configs were generated by an LLM-seeded autotuner using Claude Opus 4.8. The work lives in a Red Hat vLLM fork, not upstream.
The reason this is not upstream is the actual finding, and the post states it plainly: shipping pre-tuned configs for popular models is a maintenance burden nobody wants, large config files cannot be validated in CI, and fine-grained tuning still takes hours per deployment. So the proposal is to upstream the kernels and push autotuning onto whoever is serving the model — which is a reasonable trade for a serving provider and a non-starter for anyone running vLLM casually. The 32-token dispatch threshold is the detail to copy regardless of Helion: the gains are concentrated in small-batch decode where library kernels are least tuned, and restricting a risky optimisation to the shape range that benefits is how you get the win without the regression. All numbers are the authors’ own on one GPU generation, and the Blackwell path through Helion’s CuteDSL backend is explicitly future work.
Developer Tools #
Cloudflare adds a web search API to AI Gateway with Exa, Linkup and Ceramic.ai #
Cloudflare
AI Gateway can now inject live web results into model inference calls through three launch partners — Ceramic.ai, Exa and Linkup — reachable by REST at api.cloudflare.com/client/v4/accounts/{ID}/ai/websearch/, by a one-line Workers binding, and soon as server-side tools inside the control plane. Billing runs through AI Gateway credits at the partners’ list prices with no Cloudflare markup, bring-your-own-key is supported, and zero-data-retention partners are identified as such. All three partners committed to Cloudflare’s Verified Bots requirements and to including source links in results. No latency or quality figures were published.
The crawling commitment is the distinguishing term, not the API. Cloudflare sits on both sides of this transaction — it sells bot management to publishers and now resells search to agent builders — so requiring partners to crawl as verified bots and to return source links is the company making its two businesses consistent rather than a courtesy. For a builder the practical consequence is narrower: search is one of the two or three calls every agent makes, and consolidating it behind the same gateway that already holds the model credentials removes a key and a vendor relationship. Charging at partner list price with no markup means the margin is elsewhere, which is worth remembering when the pricing changes. No published latency is the gap, since a search call sits in the critical path of every grounded response.
Research & Papers #
Google deploys federated learning inside TEEs with policies published to a public transparency log #
Google Research
The system has client devices encrypt their training examples and pre-authorise an access policy; a key management cluster running RAFT consensus releases decryption keys only to workloads whose measurement matches a published policy; data-processing trusted execution environments then run the training program and delegate to worker TEEs. The access policies are published to Rekor, a public transparency log, so an outside party can check which workloads were ever permitted to touch the data. Only metrics and differentially private model weights leave the enclave, the training logic is reproducibly built from open source, and encrypted uploads are decryptable only inside a TEE for a limited window after upload. Gboard has shipped English and Japanese next-word prediction models trained this way, with training time falling from one to two months to substantially less through server-side parallelisation.
The substantive change is where the trust sits. Classic federated learning keeps data on the device and asks you to trust that the server aggregates honestly; this uploads encrypted data to the server and makes the server’s behaviour externally checkable instead, which is a better bargain if and only if the attestation chain holds. Publishing policies to Rekor is the part that is genuinely novel in a production ML system — it borrows the supply-chain transparency pattern and applies it to data access, so the claim is auditable after the fact rather than asserted in a privacy policy. Two things to keep in view: the guarantee is now contingent on TEE integrity, which has a long history of side-channel breaks, and the blog post gives no epsilon, so the differential-privacy half of the claim is unquantified here.
Putting KV cache on high-bandwidth flash cuts LLM serving completion time 36-87%, with write lifetime extended to 14.82 years #
UC Berkeley and FuriosaAI / Semiconductor Engineering
The paper characterises High Bandwidth Flash as a capacity tier between HBM and host memory for LLM serving, where weights and KV cache have outgrown what HBM can hold. The authors propose an HBM-HBF-host hierarchy with buffered cache-aware scheduling and report 36.1% to 87.0% faster completion than HBM-only systems and up to 55.8% lower energy, though some light workloads consumed more. The scheduling contribution is aimed at flash’s wear limit: it raises estimated HBF write lifetime from 4.77 years to 14.82.
Write endurance is the reason this is a scheduling paper rather than a hardware one. A KV cache is the worst possible flash workload — written once, read a few times, discarded — and 4.77 years of estimated life means the naive design wears out the device inside a normal server refresh cycle, which is what has kept flash out of the inference memory hierarchy despite the obvious capacity argument. Tripling that to 14.82 years is what converts the idea from a demo into something an operator could depreciate. The honest caveat is in the authors’ own numbers: light workloads got worse on energy, so the win depends on contexts long enough that HBM was the binding constraint, and this is a characterisation study with a proposed scheduler, not a deployed system.
Funding & Business #
Kevin Mandia’s Armadin raises $255.5 million at over $2.5 billion, seven months out of stealth #
Help Net Security / FinTech Global / TechFundingNews
Andreessen Horowitz and Accel co-led a $255.5 million Series B valuing Armadin above $2.5 billion, seven months after the company emerged from stealth, taking total funding to $445 million. Bain Capital Ventures and Redpoint are new investors; 8VC, Ballistic Ventures, Google Ventures, In-Q-Tel, Kleiner Perkins and Menlo Ventures all re-upped. The company deploys automated agent swarms that run continuous vulnerability testing for enterprise and government customers, positioning autonomous offensive security as the necessary response to attackers who now have the same tooling. Mandia founded Mandiant in 2004 and sold it to Google for $5.4 billion in 2022.
In-Q-Tel on the cap table alongside a $2.5 billion mark seven months out of stealth is the signal, since the CIA’s venture arm is not a growth investor and its presence usually means a government customer exists or is expected. The underlying thesis — that automated offence is the only way to keep pace with automated offence — is the most plausible version of the agent-security pitch, because continuous testing is one of the few agent applications where a wrong answer is cheap: a false positive costs an engineer an hour, which is the opposite of the error economics in every agentic workflow that writes to production. No revenue was disclosed, and $445 million raised against a product category that barely existed in March is a valuation resting on the founder’s track record and the market’s fear rather than on numbers a reader can inspect.
Anthropic commits $100 million to train 10,000 deployment engineers by the end of 2027 #
Anthropic
The Claude Frontier Academy is structured on a medical-residency model. Engineers start with a multi-day in-person programme built around simulated enterprise deployments, running from use-case selection through security review; those who pass an assessment earn a Claude Resident Engineer badge and move into a 12-week residency leading real Claude projects inside their own organisation, supported by Anthropic engineers, ending in a Frontier Deployed Engineer badge. First cohorts run in San Francisco, New York and London, with founding participants including Accenture, Bain, Capgemini, Commonwealth Bank of Australia, Deloitte, McKinsey, Morgan Stanley and Novo Nordisk. First FDE badges are expected in early 2027. The target is 10,000 engineers against a $100 million commitment.
$10,000 per engineer is a rounding error against the enterprise contracts this is built to unblock, which is what makes the programme legible: the constraint on frontier-model revenue is not capability or price, it is the shortage of people who can take a model from pilot to production inside a bank. Running the residency at the participant’s own employer rather than at Anthropic is the structurally interesting choice, since it means the training artefact is a deployed system rather than a certificate. The partner list says what the programme really is — six of the eight founding participants are consultancies or banks, so this is a channel strategy with a curriculum attached, and the badge is a credential whose value depends entirely on Anthropic continuing to honour it.
Sean Parker rebuilds Stability AI around music, with Sony, Warner and Universal as investors and licensors #
TechCrunch
Parker, who joined two years ago as part of an $80 million rescue financing, is now steering Stability toward becoming a tool vendor for music professionals, with Prem Akkaraju as CEO. In late August 2026 the company raised $76 million from Sony, Warner and Universal, who also licensed their catalogues for training as part of the deal. Stability has released three new audio models and music-editing software that generates instrumental tracks or snippets from text prompts, with a forthcoming feature that lets a user hum a melody or beatbox a drum pattern to steer generation.
Licensing the catalogue as part of the investment is the whole strategy in one clause: it converts the training-data question from a litigation risk into a term sheet, and it is the first time the three majors have taken equity in a generative music company rather than suing one. What that buys Stability is the only durable moat available in this category, since the models are not the hard part. The obvious counterpoint is that a tool built to the majors’ specifications is a tool constrained by their interests, and the stated positioning — for music professionals, not consumers — is what that constraint looks like in product form. Parker arriving at the labels’ table with their blessing twenty-five years after Napster is the sort of symmetry that writes itself; the substance is that the labels decided licensing beat litigating.
Other #
Wagtail spent a month trying to code on one efficient model, missed the target, and burned 3.5x its energy budget #
Wagtail
The Wagtail team set out in September 2026 to do its AI-assisted development work on GLM 5.3 Flash alone, picking it for a 1M context window, vision support and multi-provider availability. They used about 2 billion tokens over the month, roughly half of which went to other models, missing the stated goal of keeping more than 50% on the target model. The first half-month of GLM work cost $68. Energy consumption came to 35 kWh against a planned 10 kWh. They separately measured DeepSeek V4.1 Flash at 95% accuracy on their task set at $0.09 per task. Their conclusions: single-model routine development is viable, but they need better measurement, a separate budget for experimentation, and better multi-agent technique; October’s target is to keep the majority of inference on flash-tier models.
Published negative results from a team’s own tooling budget are rare enough to be worth reading, and the useful finding is not about GLM. It is that a team that deliberately set out to track its model usage still could not keep half of it on the model it chose, which says the leakage is structural — every agent harness, IDE integration and default dropdown has its own model setting, and the aggregate is nobody’s dashboard. The 35 kWh against 10 kWh is the same failure in a different unit. The $68 figure covers half a month and only the GLM share, so it is not a monthly cost for this workload, and the DeepSeek accuracy number is on their own unpublished task set.
Threads to Watch #
Agent reach is becoming an operating-system question, and the two platforms are moving in opposite directions. Apple is adding friction to Full Disk Access and naming autonomous agents as the reason, after a columnist found Meta’s Muse reading his private messages. On the same day, Meta published Apache 2.0 firmware and SDKs that put Muse on an ESP32 or a Raspberry Pi, where a “custom command for system administration” is a documented feature and there is no permission model at all. Both are correct responses to the same fact — that an agent’s usefulness is proportional to its access — and they diverge on who should hold the dial. What neither addresses is that Apple’s change is consent friction rather than scoping: a user who grants Full Disk Access still grants everything, which is the authority problem the past fortnight of agent-security research keeps arriving at from the other side.
Local inference is being attacked at the memory wall from four directions at once. ds4 quantizes MoE routed experts to 2 bits asymmetrically and spills the KV cache to SSD; the Berkeley and FuriosaAI paper puts the KV cache on high-bandwidth flash and spends its contribution on write endurance; NVIDIA halves DGX Spark’s unified memory to hit $4,999; the Helion work squeezes over 10% end-to-end out of vLLM by autotuning small-batch decode kernels per shape. None of these is about FLOPs. The shared premise is that capacity and bandwidth bound serving now, and the shared gap is evaluation — ds4 publishes no quality numbers for its 2-bit experts, NVIDIA publishes a parameter ceiling and no decode throughput, and the HBF result degrades on light workloads. The engineering is converging faster than the measurement.
Money is flowing to the layer around the model rather than the model. Armadin raised $255.5 million at $2.5 billion for agent swarms that test other people’s systems; Anthropic committed $100 million not to training a model but to training 10,000 humans who can deploy one; Stability’s $76 million came with licensed catalogues attached, which is the asset the model cannot supply. Each is a bet that capability is no longer the binding constraint — deployment skill, verification capacity and rights are. Circuit Breaker Labs is the small version of the same bet, selling simulated users to people who already have a model and cannot tell whether it is safe in the hands of real ones.
Sources Unavailable Today #
These sources could not be fetched today. Links point to their homepages so you can check them directly.
- Google DeepMind — parse_error