DeepSeek opens DeepInfer, ChatGPT adds locked chats, and Perplexity runs fully locally on Mac
DeepSeek opens DeepInfer, ChatGPT introduces locked conversations, and Perplexity Computer runs fully locally on Mac. Developer updates include TorchTPU and persistent agent memory. Research coverage spans Claude’s scorpion-toxin binders, Stanford’s medical workflow benchmarks, and OpenAI’s acquisition of Orchestra Bio.
01
Models and Personal Assistants
4 stories
2026-10-20DeepSeek
DeepSeek releases DeepInfer, its inference engine for V4.1-Flash
DeepSeek has released DeepInfer under Apache 2.0, opening the core inference engine behind its production services. The initial release supports DeepSeek-V4.1-Flash on NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs, with speculative decoding, asynchronous scheduling, and structured outputs.
The release provides offline inference. A complete online service still needs a KV cache system, prefill/decode disaggregation, and load balancing. The accompanying deepseek-recipe libraries convert request formats, encode and decode tokens, and parse outputs for Chat Completions, Responses, and Messages interfaces.
2026-10-20OpenAI
ChatGPT adds Locked Chats with PIN and biometric access
ChatGPT is gradually rolling out Locked Chats, which hide selected conversations from regular history, pinned chats, and search. Access requires a four-digit account PIN, with Face ID, Touch ID, or device unlock available on supported mobile devices. The feature covers Free, Go, Plus, Pro, Business, Enterprise, and Edu plans.
New messages in locked chats do not update regular chat memory, but information remembered before locking is not automatically removed. Locking does not provide end-to-end encryption or change retention and model-training settings. Resetting the PIN permanently deletes locked conversations and related content protected by the previous PIN.
2026-10-20Perplexity
Perplexity Computer runs fully locally on Mac, with 36GB or more recommended
Perplexity has added a fully local mode to Computer on Mac, keeping the model, planning, and execution on the device. With permission, it can work across local files and apps such as Notes and Mail. Initial model choices are Qwen 3.6 and PPLX Pro, a post-trained Qwen variant.
The feature is available to Pro, Max, and Enterprise subscribers on supported Apple silicon Macs; Perplexity recommends at least 36GB of memory. Unlike Hybrid Compute, a fully local task needs no cloud inference. Users can still authorize cloud escalation for current information, browser use, or connected apps.
2026-10-19TechCrunch
Wabi opens to all users with shared assistant threads and custom apps
Replika founder Eugenia Kuyda has opened Wabi to all users. The assistant organizes tasks and long-term goals into messenger-style threads, each of which can contain custom apps for activities such as health tracking, family outings, or restaurant suggestions. Threads can also become shared group conversations.
Kuyda told TechCrunch that connecting email lets Wabi create initial threads and apps automatically. Shared information is scoped to individual chats, where members can see requests for personal information. Wabi currently runs in the cloud and is free; the company has not settled on a business model.
02
Developer Tools and Agent Engineering
3 stories
2026-10-20PyTorch / Google
Google and Meta launch TorchTPU for native PyTorch training and inference
Google and Meta have introduced open-source TorchTPU, making TPU a native PyTorch device. It supports eager execution, torch.compile, and distributed operations. The announcement illustrates basic migration by changing the device target from CUDA to TPU and says major customers already use it for training and inference.
Google also released a TPU fork of TorchTitan and an optimization guide, plus native vLLM and SGLang serving backends. The serving backends retain existing schedulers and APIs, supporting prefix caching, tool calling, and structured output. Google says SGLang-TorchTPU already runs in production at Meta.
2026-10-19AWS
SageMaker Studio simplifies signup and starts new users with $100 in credits
New AWS users can sign up with Google, GitHub, Apple, Amazon, or email and enter a preconfigured SageMaker Studio environment. AWS says setup takes about 20 seconds, and most new customers do not need a credit card or the previous full AWS signup process.
Accounts receive $100 in credits and can earn another $100 for training, customization, notebooks, and endpoint deployments. The free period ends after six months or when credits run out, whichever comes first, at which point resource usage stops. Upgrading requires payment details; existing accounts keep their current sign-in process.
2026-10-20Databricks
Databricks introduces MemCurrent for persistent agent memory
Databricks has detailed MemCurrent, the persistent memory architecture behind Agent Bricks Memory and Genie. It stores raw session traces separately from distilled knowledge. Agents can update memories immediately, while a background process merges duplicates, resolves conflicts, and splits oversized files.
On an internal benchmark with 114 cases from 29 users, task pass rate rose from 22.3% without memory to 68.4%, with 31% fewer tokens and 26% less runtime. The test used Claude Opus 4.6 for tasks and GPT 5.6 Sol for memory, so these are results for the company’s specific workflow.
03
Science and Healthcare
3 stories
2026-10-20Anthropic
Claude-designed scorpion-toxin binders yield 17 qualifying results from 211 tests
Anthropic let an internal research model work largely autonomously in Claude Science for 64 hours, generating 318,484 protein designs against five scorpion neurotoxins. It selected 180 designs, and the human team added 31 with its help. Partner laboratories tested all 211; 17 met the specified binding criteria.
The successful designs bind to four of the toxins. Humans supplied an initial prompt and four follow-ups, and the designs and experimental data are public. The team has not established venom neutralization in a cellular environment, so the results are candidate binders, not a demonstrated antivenom.
2026-10-19Stanford HAI
Stanford medical benchmarks find leading agents completing fewer than half of full workflows
Stanford HAI has reviewed its progression from MedHELM to workflow-based medical evaluations. PhysicianBench tests 100 multi-step tasks across 21 specialties in an electronic-record sandbox. Among 12 models, GPT-5.5 led with 46% completion on a single attempt; just 28% of tasks succeeded in all three runs.
HealthAdminBench uses synthetic data across record, insurance, and fax systems for 135 administrative tasks. Claude Opus 4.6 completed 36.3% end to end. The environments, model sets, and grading differ, so the scores are not directly comparable. Both expose failures to carry information across documents and systems.
2026-10-20OpenAI
OpenAI acquires Orchestra Bio to connect drug-program coordination with Rosalind
OpenAI has acquired Orchestra Bio and is bringing its team and technology together with Rosalind’s life-science models and tools. Orchestra helps biopharma staff and agents track studies, project progress, and dependencies across teams, adding coordination capabilities to the path from scientific discovery to drug development.
OpenAI gives the example of a failed manufacturing batch delaying a clinical trial. Orchestra agents can identify affected activities and timelines, gather information, and compare responses for scientists and program leaders to consider. The announcement does not disclose financial terms.
04
Safety and Governance
2 stories
2026-10-19Redact
Redact finds restrictive hidden AI instructions on WhatsApp help pages; notes are removed
Redact reports that WhatsApp help pages contained hidden text directing its support chatbot not to recommend data-protection regulators. When users mentioned lawyers or litigation, the notes also instructed it to avoid citing privacy rules such as GDPR, LGPD, and CCPA. Researchers found 102 hidden blocks across 80 pages.
The instructions differed from information in the user-facing privacy policy. Redact says the notes were still present on October 15 but had all been removed by its October 18 check; the help pages remained online. Its report also cites WhatsApp’s confirmation to WIRED that the notes were intended for its own support chatbot.
2026-10-20METR
METR broadly agrees with a low-risk judgment but questions Anthropic’s supporting methods
METR has reviewed the high-stakes misalignment section of Anthropic’s August 2026 risk report. It broadly agrees that catastrophic misalignment risk is low for the covered models, primarily because their capabilities remain limited, rather than because safeguards and alignment training have been decisively validated.
METR doubts whether the methods would reliably detect higher risk and says recent incidents weaken some arguments. This review relied only on the internal report and predates the newer independent-evaluator arrangement, without its broader access. Anthropic retained publication-review rights but made no redactions to the review.