Sections
00:00 What we're covering today
00:34 TSMC reports record profit and keeps its sixty to sixty-four billion U S dollar spending plan
02:09 Meta releases weights for a model with 816 billion parameters
03:14 GitHub releases AI scans, but they cannot block pull request merges
04:24 Anthropic gives six M E T R employees internal access and lets M E T R publish without approval
05:36 Nine hundred fifty-two thousand eight clinicians asked OpenEvidence two hundred seventy-one million questions
06:52 Visit Hot Tea

Disclosure
Narration uses an AI-generated voice.

Transcript
Welcome to Hot Tea for Friday, October sixteenth, twenty twenty-six. This briefing uses an A I-generated voice.

TSMC just reported record profit, but it didn't raise its capital budget as the cost of overseas factories comes into view. We'll also look at Meta's terms for larger model-hosting businesses and GitHub's new advisory security scans.

Then we'll turn to M E T R's access inside Anthropic, and what millions of clinician questions can and can't tell us about medical A I.

TSMC reports record profit and keeps its sixty to sixty-four billion U S dollar spending plan. TSMC announced its third-quarter results on October fifteenth, for the quarter that ended September thirtieth. Profit reached seven hundred sixty-three point zero six billion New Taiwan dollars, up sixty-eight point seven percent from a year earlier. Revenue reached forty-six point seven two billion U S dollars, up forty-one point two percent in that currency. Management raised its full-year revenue growth forecast to about forty-four percent in U S dollars. But it kept the company's twenty twenty-six capital budget at sixty to sixty-four billion U S dollars. That budget didn't rise in this announcement. Demand includes advanced chips for A I accelerators, processors and networking equipment. Still, these results cover all of TSMC, and the company didn't report a separate A I revenue figure. New factories also bring costs. TSMC expects its overseas factories to reduce its overall gross margin by about two percentage points in twenty twenty-six. It expects that effect to reach three to four percentage points over the next few years. Those estimates are forecasts, and so is its revenue outlook. For the fourth quarter, TSMC forecasts revenue of fifty-two point three to fifty-four point three billion U S dollars. Its next results will show whether revenue reached that range, and how much it cost to bring the new factories online. TSMC can produce chips profitably without proving that each customer's A I service makes money.

Meta releases weights for a model with 816 billion parameters. Muse Spark Open uses forty-two billion parameters for each token, and its context window holds about one million tokens. Meta released the weights on October fifteenth. Several providers also offer hosted access. The Muse Model License has a separate rule for businesses that offer models as a service. The trigger is more than twenty million U S dollars in gross revenue, including affiliates, during any consecutive twelve-month period. A business that crosses that threshold must notify Meta. It also needs a separate written agreement before offering this model or its derivatives through that service. That condition doesn't apply to internal use, research or commercial activity that doesn't offer the model as a service. Product features outside that model-hosting business are also exempt. Meta supplied the performance and safety comparisons. Independent tests still need to measure running costs and agent performance. Any company planning to host the model for customers also needs to check those license terms.

GitHub releases A I scans, but they cannot block pull request merges. GitHub's October fifteenth release checks languages and frameworks that Code Q L doesn't cover today. A I Scan findings are advisory, so they can't fail the code-scanning check or block merges through rulesets. The tool also excludes pull requests from forks and Dependabot. GitHub separately released agentic autofix. It proposes repairs, reruns Code Q L and opens a draft pull request for a person to review. Even when that recheck succeeds, it doesn't prove that every security problem is fixed. A I Scan needs code scanning and the relevant setting enabled, and it uses Code Security A I Credits. Private and internal organization repositories need Code Security or Advanced Security. Agentic autofix has separate eligibility rules on GitHub Enterprise Cloud. It needs administrator approval and a defined Copilot billing method. It also uses A I Credits and Actions minutes. GitHub described both tools. Their use in real repositories will show how often they miss flaws, raise false alarms or propose repairs that developers accept. It will also show what that work costs.

Anthropic gives six M E T R employees internal access and lets M E T R publish without approval. Anthropic announced the arrangement on October fifteenth. The six employees have access to company hardware and office space. They also have internal access similar to what Anthropic gives its risk teams. M E T R can publish its assessments without Anthropic's approval. Anthropic can still require specific information to be removed within narrow, predefined categories. M E T R can describe and dispute those removals in public, but it can't reveal the removed information. Anthropic doesn't routinely share customer data, privileged information or details of its model architecture. The company granted this access voluntarily. It isn't a government inspection, and it isn't a finding that the models are safe. Other evaluator partnerships offer less access and narrower publication rights. M E T R's October fourteenth policy says readers need to know when companies limit access or remove important details. The harder test will come if a finding could hurt Anthropic. The useful signs will be whether M E T R describes withheld evidence or disputes the removal of information.

Nine hundred fifty-two thousand eight clinicians asked OpenEvidence two hundred seventy-one million questions. The JAMA study appeared on October fifteenth. It examined how U S clinicians used OpenEvidence from January first, two thousand and twenty-four, through June thirtieth, two thousand and twenty-six. Five hundred fifty-six thousand two hundred were physicians. The others included nurse practitioners, physician assistants and other licensed clinicians. Monthly active users reached seven hundred three thousand two hundred sixty-five in June two thousand and twenty-six. A clinician counted as active after asking at least one question that month. Each follow-up counted as another query, so the total isn't a patient count. This was an observational study of one platform. It didn't establish whether the answers were accurate or helped patients. The researchers also couldn't tell whether clinicians asked questions during patient care, while preparing for it or as part of general learning. OpenEvidence supplied the data and helped analyze it, and several authors worked for the company. JAMA's accompanying editorial also warns that citations don't validate generated answers. The next evidence needs to connect answers about treatment and medication with clinical decisions and patient outcomes. Repeated use alone doesn't establish better care.

That is the signal before the noise. This briefing was produced from Hot Tea's verified daily edition. For the complete briefing and every source link, visit Hot Tea dot A I.