Morning Brief · October 15th 2026

Morning Brief, October 15th 2026

TSMC beat its revenue forecast and projected higher sales next quarter. Anthropic reported $16.8 billion in quarterly sales to investors. Eight states sued OpenAI. A coordinated AI system solved nine of ten unpublished mathematics problems.

18 min read·Editorial by Vera Lindqvist

TSMC reports $46.72 billion in quarterly revenue and guides higher

TSMC reported third-quarter revenue of $46.72 billion on October 15th, above its earlier $44.6 billion to $45.8 billion . Gross margin reached 66.4%. Its fourth-quarter revenue forecast is $52.3 billion to $54.3 billion, with between 65% and 67%.

The fourth-quarter figures are management guidance. The distinction matters when chip demand becomes evidence for an entire industry's economics. TSMC's customers are ordering enough manufacturing capacity to support another increase in sales. Whether their own AI products earn an adequate return is a separate calculation, and these margins do not answer it. SourcesA

First Proof finds nine correct solutions from a coordinated AI system

First Proof's third evaluation gave five systems ten unpublished research mathematics problems, with no human intervention during their attempts. Its October 14th report says ProofCouncil solved nine correctly; GPT-6 Astra alone solved seven. Thirty reviewers assessed correctness, exposition and attribution without knowing which system produced each submission.

The organizer reports costs of $1,888 for ProofCouncil and $36 for Astra. The Kimi K3 system submitted no correct solutions. These are results for particular systems and workflows, not a controlled comparison of model families. Correctness also outran citation quality: Astra received adequate-attribution judgments on four problems. A useful proof still owes its reader the work it borrowed.

Correct solutions on First Proof's third batch

Three of five evaluated systems shown. Workflows and spending differ; organizer-reported results.

Anthropic tells investors quarterly revenue reached $16.8 billion

Anthropic told investors that third-quarter revenue reached $16.8 billion, Reuters reported on October 14th, citing two people who attended its investor day. Reuters puts second-quarter revenue at $11.6 billion, making the increase about 45%. The company also reported profitability on an adjusted basis. FT reporting relayed by El País gives the prior quarter as $11.5 billion and growth as 46%; the comparison here uses Reuters' attendee-sourced $11.6 billion without averaging them.

The revenue comparison is stronger evidence than an annualized sales pace: it describes consecutive quarters. The profit claim needs a different reading. Adjusted earnings can exclude costs that shareholders still bear, and Reuters obtained the figures from meeting attendees; audited statements would allow outsiders to check them. Investors have a new operating number to assess. They do not yet have a completed or enough public detail to reconstruct the profit calculation. SourcesB B

Eight states sue OpenAI over ChatGPT safety claims

A bipartisan group of eight states sued OpenAI in separate state courts, Reuters reported on October 14th. Iowa, Kansas, Arizona, Arkansas, Idaho, Montana, Nebraska and West Virginia allege that the company misled users about ChatGPT's safety and that the product can reinforce delusions and encourage harmful behavior.

The complaints invoke consumer-protection and public-nuisance law. Requested remedies include product changes, financial penalties and limits on self-improving models without assurances about human values. The courts have yet to decide these allegations or grant the requested restrictions. Their immediate significance is procedural: several state courts now face claims about the safety of the same product. Reuters said OpenAI had not immediately responded to its request for comment. SourcesB

Updated

Ling 3.1 Flash adds downloadable MIT-licensed weights

InclusionAI has published downloadable Ling 3.1 Flash under the , extending the model beyond its previously available hosted access. The specifies 560 billion total parameters, 25 billion active parameters and a context window of one million tokens.

Its architecture combines linear attention with a smaller number of full-attention layers. That is a design for reducing the cost of long inputs, not a guarantee that every long document will be understood correctly. The practical change is control: operators can inspect the release and plan their own deployment instead of depending exclusively on a hosted . Hardware requirements and the vendor's claims still need workload-specific testing. SourcesA

Bank of America keeps a thousand contractors on vulnerability repairs

Bank of America chief executive Brian Moynihan said roughly 1,000 contractors hired in the second quarter are still working on vulnerabilities exposed by AI, including Mythos, according to The Next Web's account of his October 14th remarks. They work alongside about 3,000 internal staff on the effort.

Finding defects does not repair a production bank. Changes have to survive testing, dependencies and release controls. That makes labor a deployment cost even when discovery gets cheaper. Moynihan also discussed spending of a couple hundred million dollars or more across anti-money-laundering and Mythos work; that combined figure cannot honestly be assigned entirely to AI security. The staffing figures come from management and were not independently audited in the report. SourcesB

Updated

Amap investigation adds logs and a disputed explanation of Tencent-linked agents

Swarmchasers published a fuller investigation on October 14th, documenting more than 2,600 attempts against Alibaba's Amap service between September 28th and October 8th. The researchers describe using Tencent Cloud infrastructure and a website-scanning service to bypass restrictions. Most attempts failed; some retrieved data.

The new report also carries a competing explanation. A source close to Tencent said an evaluation team used third-party models, that the activity had ended, and that an internal review found no coordinated targeting or breach. That is not an official public response. Infrastructure traces do not independently establish the model or the person directing it. The researchers found no evidence of real-time coordination between the agents. The logs support scrutiny; the attribution remains contested. SourcesA

OpenAI consolidates compute leadership as Sachin Katti prepares to leave

OpenAI is expanding Uday Ruddarraju's responsibilities to cover both compute and infrastructure, according to PYMNTS' October 14th account of Bloomberg reporting. Sachin Katti is due to leave on November 15th and remain an adviser. Ruddarraju previously worked at xAI and joined OpenAI in 2025.

Consolidating those responsibilities puts hardware acquisition and the systems that use it under one executive. That can shorten decisions about where to spend scarce capacity, but an organization chart does not establish better utilization or delivery. The reported departure is prospective. The useful follow-up will be whether OpenAI meets its deployment commitments under the new structure, under the expanded remit. SourcesB

Google opens Deep Review for technical documents

Google Research opened Deep Review on October 14th, offering two free technical-document reviews per week, with higher usage available through an key. The initial interface accepts PDFs and plain text. Multiple-document support is described as future work.

The system assigns parallel checks, adjusts its effort and assembles a report grounded in outside sources. It grew out of opt-in pilots that were kept separate from official conference reviewing. A useful deployment constraint sits beside the capability claim: Google says uploads are not used for training, while the interface describes a one-year retention period. Authors handling confidential material should assess both statements. Automated criticism can improve a draft without being qualified to accept a paper. SourcesA

Designed enzymes cut proteins in cells, with efficiency still a limit

A Nature paper published October 14th reports the design of cysteine proteases, enzymes that cut proteins using a particular chemical reaction. Thirteen of 69 tested designs were active. Six experimentally determined structures closely matched the intended designs, and the researchers demonstrated cleavage of targets in mammalian cells.

Those measurements connect generated structures to physical function. They also expose the remaining engineering problem: the designs are less active than the best natural enzymes, with the paper identifying imperfect positioning and preparation of the reacting components. This is experimental evidence for designing a molecular tool. It does not establish a safe treatment, delivery into patients or a clinically useful dose. The next improvement has to survive another physical experiment. SourcesA

Grok Voice 3 arrives through the existing voice API

xAI announced Grok Voice 3 on October 14th and said the existing latest-model alias would move to the new version automatically. The release emphasizes conversational , instruction following and handling support calls.

An automatic upgrade is a deployment event for anyone who built tests around the previous model's behavior. The sensible action is to difficult calls and verify escalation, interruptions and tool use before expanding traffic. xAI's account includes results from Starlink support calls, but a vendor-selected deployment does not establish the same resolution rate for another company's customers. The release provides a new service to test. It does not remove the need to measure what callers actually achieved. SourcesA

JetBrains releases Mellum 2.1 for coding agents

JetBrains released Mellum 2.1 under , keeping a 12-billion-parameter architecture with 2.5 billion active during . The release announcement describes reinforcement learning in millions of sandboxed tasks built around real repositories.

The attraction is a coding model that an operator can host and adapt under a permissive license. Its speed and task-completion claims remain vendor measurements, and a smaller count does not specify the memory needed to hold the whole model. JetBrains lists additional local formats and serving improvements as forthcoming. Teams should distinguish downloadable artifacts from the compatibility work promised next, then measure whether the model completes repository tasks. SourcesA

GitHub replaces generative secret classification with ModernBERT

GitHub described a secret-detection pipeline on October 13th that combines pattern matching with a fine-tuned classifier. Pattern matching finds candidate strings; the model uses surrounding context to judge whether they are secrets. GitHub reports more than 98% lower serving costs and an 11.5% improvement in owner-weighted acceptance.

Acceptance is a product metric, not the same thing as detection accuracy. The architectural result is still useful: a bounded classification task does not need a model to write an answer. Developers get value when the scanner finds they should revoke without burying them in false alarms. Cost savings matter if the system preserves that balance, including on secrets unlike its training examples. SourcesA

Earth AI agents gain private-data workflows and new evaluation tasks

Google Research is expanding experimental access to Ask Google Earth and an agent API, according to its October 14th announcement. The system combines geospatial information with private tables and geographic files, develops a plan, and executes spatial operations outside the language model's .

Keeping the map computation in a dedicated workspace is the engineering detail worth testing. It avoids treating every geographic operation as more text for the model to remember. Google also released evaluation tasks, while describing its workflow-based automated judging as an initial approach. A plausible sequence of operations can still end in the wrong geographic conclusion. Access remains experimental, and the announcement is not evidence that every generated map has been independently checked. SourcesA

Labelbox combines equity and credit in a financing above $100 million

Labelbox announced financing of more than $100 million at a $1.25 billion , led by Infinitum Partners. Its October 14th announcement includes both equity and a . The total therefore should not be described as money raised entirely by selling shares.

The company is funding work on reinforcement-learning environments through Horizon and managed agents through Recursion. That gives the financing a product thesis: customers need places to train agents and a service to run them. It also leaves a financial question. Without a disclosed split between equity and credit, the headline amount does not tell readers either the or the debt available to draw. Those are different obligations. SourcesA

Hassabis calls AI hacking incidents a warning for governments

Demis Hassabis called recent AI hacking incidents early warnings at an event in Dublin, Reuters reported on October 14th. The Google DeepMind chief argued that larger incidents could follow and called for cooperation between industry and governments. Microsoft president Brad Smith also pressed for binding legislation.

The consequential question is who gets to certify a system before it is deployed, and who can stop it after a failure. Hassabis favored risk-proportionate oversight that can change with capabilities. Those remarks add a laboratory leader's support to the regulatory debate, but they do not create a certification body, a binding threshold or an enforcement power. Each of those would require a separate public decision. SourcesB

Google offers European scientists model access and compute credits

Google announced support for up to 200 European scientists, including $3 million in compute credits and access to scientific AI systems. The October 14th announcement names tools including AlphaFold 3, AlphaGenome, AlphaEvolve and WeatherNext and connects the program to Europe's RAISE research initiative.

The practical benefit is access to systems that a small research group may struggle to run or obtain on its own. Access is also a dependency: credits expire, hosted interfaces change and model availability is controlled by the provider. The announcement does not promise downloadable weights for every listed system. Researchers should record the versions and conditions needed to reproduce an experiment before committing a project to the offer. SourcesA

DOE selects four laboratory teams for robotic science testbeds

The US Department of Energy announced four national-laboratory-led selections on October 8th for work linking robotics and automated science. Argonne, Brookhaven, Oak Ridge and SLAC lead projects covering modular experimentation, scientific validity, interoperable records and work at photon facilities.

DOE describes a $30 million total, with $2 million in fiscal-year 2026 funding and the remainder subject to future appropriations. Selection is not the same as all of that money being available now. The research problem is concrete: an agent needs a record of what an instrument actually did, including failures, if its next experiment is to mean anything. These testbeds aim to make that connection inspectable across laboratory systems. SourcesA

Leash makes measured experiments replayable for scientific agents

Leash introduced Minos on October 14th, using previously measured chemical-binding data to let agents choose experiments and receive real measurements without running a fresh laboratory assay on every turn. The setup limits an agent to a budget of 100 molecules, with up to five choices at once and no initial successful molecule supplied.

Replay makes repeated comparisons cheap enough to study how an agent allocates experiments. It also fixes the boundary of the claim: the agent chooses from measured chemical space. Success there does not establish discovery of a new molecule outside it, let alone a treatment. Leash has released research material and some evaluation packages, giving other researchers a way to examine that distinction. SourcesA

Editorial

VLVera Lindqvist
Capability & Systems

The expensive part is checking the answer where it will be used.

A mathematical reviewer can inspect a proof. A bank has to patch a running service. A laboratory must measure the molecule. These checks operate on different clocks, and the price of generating another candidate says little about the price of accepting it.

The useful engineering move is to make the check cheaper without changing what counts as success. GitHub's classifier is a narrow example: decide whether a candidate string is a secret instead of asking a general model to compose a judgment. Minos takes a different route, reusing physical measurements to compare experimental choices. Neither eliminates the need to check the final system against unfamiliar cases.

My position is that deployment capacity will be constrained more by reliable checking than by candidate generation in workflows where mistakes are costly. The evidence that would change it is a prospective trial: unfamiliar tasks, predefined success criteria, measured human review time, and no increase in downstream failures as output rises. A benchmark score without that last measurement cannot settle the question. SourcesA A

Prediction Watch

Supporting evidence. Anthropic IPOs before OpenAI: reported quarterly revenue and adjusted profitability give investors more material to assess, but neither constitutes public trading. (Prediction 2026-08-06-F1). Settles August 6th 2027. SourcesB

No change. A Chinese lab a model at 2.8 trillion parameters or larger: Ling's downloadable release is below that threshold at 560 billion total parameters. (Prediction 2026-08-06-T5). Settles February 28th 2027. SourcesA

No change. A state attorney general sues OpenAI over its agents' intrusions: the eight-state litigation reported yesterday concerns consumer harm and product safety. That reporting does not establish the unauthorized-system-access allegation required by this call. (Prediction 2026-10-04-B2). Settles March 31st 2027. SourcesB

No call settled. The retrieved evidence did not establish an Anthropic listing, a qualifying larger open-weight release or the intrusion complaint specified above.

Sources

47 citations · 31 primary · 16

Afternoon Brief, October 14th 2026

Afternoon Brief, October 15th 2026