Every story, at length

Announced Tuesday at DevDay 2026 in San Francisco, Dots are persistent agents rather than single-turn chat: they continuously crawl the web, pull context from connected apps, and learn user preferences over time. They run on OpenAI's GPT-6 Astra model. Users message Dots through ChatGPT, Slack, and Microsoft Teams, with context shared across modes; Pro users can join a waitlist for iMessage and RCS on Android. The rollout begins today for the $100/month ChatGPT Pro tier, with one Dot per user initially and multiple simultaneous agents expected later.

In demos, a Dot read a user's calendar, saw they were working through dinner, and messaged two GrubHub options with pricing before the user picked one and set an order time. Another demo showed collaborative website launch. For sensitive actions—installing software or changing passwords—Dots require explicit approval, and a Custom Rules tool lets users set boundaries and require direct permission for specific tasks.

Enterprise is a stated focus. CEO Sam Altman announced specialist Dots tuned for accounting, email marketing, and legal analysis, extending the agent-coworker trend into professional workflows. OpenAI positions this against Meta's Muse agent, which recently topped smartphone app download charts after earlier tools like OpenClaw stayed confined to early adopters.

Security caveats are explicit in the source: guardrails exist, but automation still risks unintended data sharing or manipulation by online adversaries. Users should be cautious connecting existing apps, and the model-training setting on an OpenAI account carries over to Dots interactions.

Beam is a text-only mixture-of-experts model with 501 billion total parameters and 23 billion active parameters, pretrained on 23.8 trillion tokens with a 1 million token context window. Reflection describes it as trained on high-compute reinforcement learning for reasoning, coding, and agentic tasks. For scale, Z.ai's GLM-5.2 carries roughly 744 billion total parameters with 40 billion active.

Reflection's benchmark claims are unverified. The company says Beam scores on par with GLM-5.2 on advanced reasoning benchmarks and outperforms leading Western open models while using 3-4x less inference compute. It also says Beam outscores Inkling, the open model from Mira Murati's Thinking Machines Lab, on four coding tests where both report results — though Inkling is multimodal and Beam is text-only.

The startup, founded in 2024 by two former Google DeepMind researchers, has raised roughly $4.7 billion from Nvidia, Sequoia Capital, and Lightspeed Venture Partners, with a $25 billion pre-money valuation on its last round. It has also signed deals worth more than $7 billion with SpaceX and Nebius for access to Nvidia GB300 chips through 2029. Beam's weights and full technical details are due this month, with distribution through hyperscalers and neoclouds and integrations across open source libraries at launch.

Reflection is positioning Beam for enterprises, the public sector, and sovereign AI deployments. It is testing a sovereign AI factory partnership with Shinsegae Group in South Korea, and Axios reported hedge funds and trading firms are among those interested in building customized local systems on Reflection models.

Simon Willison reran a two-year-old GPT-4o experiment by Colin Frasier that tested whether models could compute sums and return answers in words across increasingly large numbers. The original GPT-4o results were poor enough that Willison was confident it had not cheated with a calculator. His rerun used Qwen3.8-27B-Q4_K_M.gguf on a DGX Spark, with the experiment orchestrated through a Codex Remote session running GPT-6 Astra.

With reasoning disabled, the model was run 30 times per number combination. Accuracy came in at 23% overall. With reasoning enabled, the run dropped to one attempt per combination because each took much longer, producing a sparse heatmap of 100% or 0% squares. The model answered correctly in 167 out of 169 attempts. Willison notes that because these were one-shot runs, a second pass would likely produce different results.

The published reasoning traces show the model working digit-by-digit from right to left, aligning columns, tracking carries, and even catching its own mistakes with lines like "Wait, let me redo this more carefully." That explicit self-correction is the mechanism behind the accuracy jump.

The 23% figure is an aggregate across all combinations tested with reasoning off; the source does not break out per-combination accuracy for that condition. The 167/169 figure is the total across all combinations with reasoning on.

The watermark is not a visible symbol. It works by subtly shaping the model's word choices, leaving a statistical pattern that is invisible to readers but detectable with a secret key. Because the pattern lives in the token sequence itself, it survives copy-paste. OpenAI said it saw no meaningful change in model performance with watermarking enabled, and that the watermark does not identify the user.

OpenAI published a technical report for textGrain, co-written with researchers from the University of Pennsylvania and Yale. The method uses a secret key to sort next-word predictions, nudging token selection in ways that accumulate into a detectable signature. Detection requires only the text and the key. The rollout covers eligible ChatGPT and Codex users on all plans in the EU over the coming weeks. API developers anywhere can enable it for select models starting today, but it is off by default, and OpenAI is not making text watermarking a global default at launch.

OpenAI's own tests show the watermark is fragile under editing. Replacing 10% of words with synonyms dropped detection from about 92% to 66%. Short passages, math answers, and translated text are also harder to detect. The company is restricting initial detector access to approved researchers and expert organizations, and cautions that a missing watermark does not prove human authorship — the text could be too short, too heavily edited, or generated by another company's model.

The move follows Anthropic's worldwide Claude watermarking announcement two months prior, which drew user backlash. OpenAI had previously built a text watermark but held off releasing it, partly over concerns that users would switch to rivals, according to a 2024 Wall Street Journal report.

Independent researcher Syed Asad Mohiuddin tested agents from Google, JP Morgan Chase, Weaviate, Rapid7, the French government's interministerial digital directorate, and the US federal government. His proof-of-concept attacks exploit trust gaps in MCP, the Model Context Protocol, which is one way AI apps and agents communicate inside an internal network. In the past five months, Google and four other organizations have acknowledged vulnerabilities that let an attacker compromise one agent and use it to spread harmful instructions to other internal agents.

The technique is a special form of prompt injection that targets a particular agent—such as one for translation or data analysis—rather than the LLM itself. Guardrails inside such special-purpose agents, if they exist at all, are often lax, and the agent will forward the instructions to other agents down the chain. Because downstream agents explicitly trust the first one, they follow the directions. This is distinct from classic prompt injection: an exploit that would have been rejected by the LLM succeeds because MCP servers store credentials for each agent and agents are built to trust every other internal agent.

In many cases, well-crafted prompts targeting the right agent lead to server-side request forgery (SSRF), a vulnerability that causes a web server to make unauthorized network requests. The broader consequence is exfiltration of database contents and sensitive business and personal information. The source describes the flaw as structural and hard to mitigate, since it arises from the trust assumptions baked into MCP deployments rather than a single implementation bug.

Anthropic has moved the Cowork execution environment off the user's device. The previous version ran model inference in the cloud but executed tool calls in an Anthropic-provided VM shipped to the local computer. That VM existed for capability, safety, and security reasons — it mapped in only the data a user explicitly added to a session. The trade-off was local resource cost: disk, battery, and performance, plus the fact that closing the laptop halted work.

The new version runs both model inference and the VM in the cloud. Each session gets its own sandbox with no shared state between sessions. When the VM needs something on the user's device, such as a file, the desktop app handles that file-access tool call. This keeps the safety boundary of explicit data mapping while removing the local VM overhead.

Felix Rieseberg at Anthropic frames the change as addressing recurring complaints: using Cowork from a phone, keeping work running when the laptop is closed, and getting the same capability without battery drain from the VM. The announcement links to a help page for further detail.

For practitioners, the architectural shift is the story: tool execution moves from a local sandbox to a per-session cloud sandbox, with the desktop client reduced to a file-access broker. That is a meaningful change in trust boundary and resource profile, not just a hosting tweak.

The release centers on tasksets: an RL environment on the Hub is a dataset repo carrying the rl-environment tag, which surfaces it under the new RL Environments filter at huggingface.co/datasets?other=rl-environment. Four frameworks are registered as dataset libraries — Harbor, Verifiers, NeMo Gym, and OpenEnv — and each framework tag adds its icon to the dataset page plus a generated loading snippet under "Use this dataset." A dataset can carry multiple framework tags; tags describe compatibility and do not convert files. The docs specify that the YAML header needs only the rl-environment tag plus every framework that can load the files.

Worked examples show the same Harbor task directories running across three loaders. Harbor's oracle agent executes a reference solution and scores it without calling a model: harbor run against harborframework/terminal-bench-2.1 with --include-task-name '*regex-log' and --agent oracle. Verifiers v1 runs the same repo in Docker with a minimal bash harness via uvx --from 'verifiers[harbor]' eval harbor, using the repo's registry.json for dataset name and version. OpenEnv 0.7.0 runs an OpenCode agent, writes rollout.json, and returns the verifier reward plus model call count; a reward of None means no verifier reward was produced. NeMo Gym covers both evaluation and RL training, with environments that collect trajectories and compute rewards.

Tagged environments already on the Hub include Harbor's BeyondSWE, Terminal-Lego, Harbor-Mix (100 tasks), and NatureBench (90 tasks); Verifiers' Reverse-Text-RL, Multi-SWE-RL-Verified (2,232 of 4,703 rows passing gold-patch validation across C, Go, Java, JavaScript, Rust, and TypeScript), and Scale-SWE-Verified (17,202 of 20,181 Python tasks); and NeMo Gym's Workplace Assistant (five databases, 26 tools, 690 tasks), Structured Outputs, CFBench, and SysBench. The stated goal is automatic framework tagging everywhere, with per-config snippets and structural detection listed as next steps.

The argument is structural: environments are tasks, tests, containers, and a reward rule — data with a runtime on top. The Hub already versions, gates, previews, and serves data, so a second registry system is unnecessary. One consequence is a single discussion tab per repo where broken tasks can be reported across frameworks, so a fix propagates to every framework on the next pull.

The core dispute is over residual value. Nvidia argues its most specialized chips can generate revenue for up to a decade, citing third-party studies showing major cloud companies extending server depreciation to five or six years, and a Barkr valuation putting the useful life of GB300 NVL72 systems at nine to 10 years. Banks underwrite GPUs on a 3-4 year depreciation schedule, according to Tony Trzcinka of Impax Asset Management, and lack the historical data to confidently underwrite long-term residual value. Andrew Chang of S&P Global Ratings said the firm takes a conservative view of chip values even though GPUs have so far proven to work well north of five years.

Nvidia's August plan, announced with Blackstone, Apollo and KKR, envisioned using chips as collateral with limited guarantees, including residual-value guarantees of no more than 25% on some deals. That is thin compared with recent precedents: CoreWeave's $8.5 billion facility received an A3 rating largely because lenders relied on contractual payments from Meta, and Broadcom backstopped more than 80% of a $35 billion financing structure for Anthropic. Nvidia itself previously provided a residual-value guarantee for SB Energy's Ohio data center project.

Reuters reports, citing unnamed banking sources, that tens of billions of dollars in loan deals now in the pipeline are likely to include stronger guarantees and contractual protections. Some structures under consideration could provide lenders with guarantees, and the initial deals will be secured by Nvidia chips and backed by customer contracts plus Nvidia's underlying guarantee. Three banking sources said Nvidia may ultimately need to provide guarantees on all deals, or have them backed by revenue streams from investment-grade customers such as major technology companies.

An Nvidia spokesperson said AI compute is "a productive, durable and fungible asset that can support long-term financing" and that financing partners independently assess each opportunity, including customer commitments, expected cash flow and residual value. Demand to finance the deals remains strong despite the guarantee concerns, the sources said.

Chakra is an AI agent that conducts interviews, observes candidates as they work in a real-world code repository, and evaluates not just the final answer but the reasoning behind it. The product went GA on Monday after beta testing with companies including Snowflake, Snorkel, and Capgemini, alongside internal use at HackerRank. The interview format gives candidates a task in a canvas with an embedded AI assistant, and Chakra uses the context of their work to ask follow-up questions — why one approach over another, or how a solution would change under a new constraint.

The company is explicitly repositioning its evaluation model. CEO Vivek Ravisankar argues that since AI lets anyone produce an artifact, the useful signal has shifted from output to process: critical thinking, judgment, and what HackerRank calls "AI fluency" — how well a candidate frames a problem for AI, judges its output, and steers it toward a solution. Ravisankar says Chakra collapses what was previously three rounds — recruiter screen, take-home assessment, and engineer follow-up — into a single interview.

On cheating, HackerRank reports the opposite of the intuitive result. Suspicious-activity flags were 70% to 80% lower in Chakra interviews than in comparable traditional HackerRank assessments, with variation by geography and seniority. The company's explanation is that giving candidates sanctioned access to AI reduces the incentive to use outside tools covertly.

Chakra is designed to score candidates rather than make the final hiring decision, which remains with humans. Ravisankar argues AI can apply a consistent rubric and is "way less biased than humans, if you tune it properly," while acknowledging that hiring is regulated and that compliance with requirements such as New York City's bias-audit rules is part of what HackerRank has had to build for. The company, launched at TechCrunch Disrupt in 2012 and Y Combinator-backed, now has more than 3,000 business customers and a community of over 30 million developers.

TikTok announced Monday that it is launching an AI shopping assistant and an in-app checkout feature that lets users buy directly from brands. The Shopping Assistant is described as a conversational AI agent that understands context and remembers user preferences and needs throughout a conversation. It provides product details, shipping information, sizing, availability, and help completing a purchase. The checkout feature lets users buy from a brand directly from the For You feed with one click.

The new features are being built in partnership with commerce platforms and payment providers including Salesforce, Shopify, Shoplazza, and Stripe. This matters because TikTok Shop, the app's in-house marketplace, has been the platform's main e-commerce engine since its U.S. launch in September 2023. These additions broaden the approach beyond that shopping hub, integrating purchasing tools into the broader app rather than routing transactions through TikTok Shop.

The strategic context is retention of shopping journeys. By pairing a shopping assistant with direct payments, TikTok turns its impulse-driven discovery feed into a place where users can get product questions answered and purchase on the spot, without leaving the platform. TikTok also appears to be positioning the assistant against outside AI tools like OpenAI's ChatGPT for product-related questions.

TikTok's Economic Impact Report, released last week, says activity on the platform helped generate $81 billion in GDP for U.S. businesses, and 54 million Americans said they had purchased a product after watching a TikTok video. eMarketer estimates TikTok Shop generated about $15.8 billion in U.S. sales in 2025, roughly 18% of all U.S. social commerce.