6 min read · 14 sources

Tesla and SpaceX have formed independent board committees to evaluate a potential merger, according to reporting from The Wall Street Journal. Elon Musk is excluded from both committees, and any transaction will face strict regulatory reviews and a full shareholder vote before closing. Consolidating orbital launch, satellite communications, manufacturing, and self-driving fleet operations under a single corporate umbrella would mark the most capital-dense corporate tie-up in tech history.

At the same time, Anthropic’s draft registration statement revealed the brutal physics of the frontier LLM compute race: an $80 billion operating loss over just six months. Between corporate consolidations and spiraling balance-sheet leverage, the cost of staying on the technological frontier is breaking past traditional venture frameworks.

Meanwhile, infrastructure engineers have two immediate operational crises: eight compromised intermediate TLS certificates leaking forged identity across the web, and an unauthenticated pair of remote code execution bugs spanning every modern version of PHP.

Eight compromised eMudhra TLS intermediates forced a global revocation cascade affecting IoT, Docker images and legacy Linux distro bundles.

Anthropic's S-1 Shows an $80B Half-Year Loss and a $750B Compute Horizon

Source: wsj.com ↗

Anthropic filed a confidential draft S-1 prospectus revealing an operating net loss exceeding $80 billion during the first half of 2026. While the company recorded rapid top-line revenue expansion, its forward liabilities highlight the severe capital intensity of frontier training runs. The filing explicitly notes that Anthropic could spend up to $750 billion over the coming years simply securing access to advanced accelerator clusters and datacenter capacity.

Beyond the raw cash burn, the prospectus lists structural risk factors around model safety liabilities and catastrophic harm. Regulators and enterprise customers are demanding indemnity for downstream model actions, and the filing concedes Anthropic lacks guaranteed mechanisms to prevent models from generating actionable harm. For operators, this means the API bills funding frontier AI remain subsidized by massive capital expenditures that will eventually force margin compression or sharp price hikes.

Tesla and SpaceX Form Committees to Explore Mega-Merger

Source: wsj.com ↗

Tesla and SpaceX boards have created dedicated special committees to explore combining the two companies. Musk sits on neither committee in an effort to satisfy fiduciary independence, as the transaction would integrate a publicly traded consumer hardware and autonomous-vehicle company with a heavily classified, privately held defense and launch contractor.

The deal faces deep antitrust and cross-agency scrutiny, particularly around national security exemptions, Starlink’s orbital data backbones, and public shareholder exposure to private rocket development risks. If approved by both special committees, the combined entity would require majority approval from Tesla’s common shareholders, forcing index funds to take on launch-market equity.

Critical PHP Remote Code Execution Vulnerabilities Trigger Out-of-Band Patches

Source: bleepingcomputer.com ↗

The PHP development team alerted systems administrators to prepare for out-of-band updates addressing two critical vulnerabilities: CVE-2026-109618 and CVE-2026-109619. The vulnerabilities carry official CVSS scores of 9.2, with security firm watchTowr tracking them as a maximum 10.0 due to unauthenticated remote code execution viability across standard default configurations.

The security advisory confirms the flaw touches PHP branches 5, 6, 7, and 8. Because technical exploitation details leaked before coordinated disclosure windows closed, malicious scanning is already underway. Teams running PHP-FPM behind Nginx, Apache, or legacy monoliths must treat these upcoming binaries as drop-everything deployments.

Eight Intermediate TLS Certificates Leaked Weak Keys at eMudhra

Source: arstechnica.com ↗

Internet infrastructure teams are scrubbing certificate bundles after eight intermediate certificates belonging to Indian CA eMudhra were discovered to have catastrophically weak private keys. The breach was proven publicly when a researcher signed an arbitrary verification message using one of eMudhra’s exposed production signing keys, validating that unauthorized parties could issue trusted TLS certificates for arbitrary hostnames.

Major desktop browsers, mobile operating systems, and root stores pushed revocation blocks for the eight intermediates via CRLite and OCSP stapling. The remaining operational headache sits in headless Linux containers, embedded IoT devices, base Docker images, and local trust stores (such as /etc/ssl/certs/ca-certificates.crt) that do not dynamically evaluate intermediate revocation chains. SREs should grep legacy base images for eMudhra roots and force rebuilds.

Anthropic and Citrini Research Use Claude to Design Scorpion Antivenom Candidates

Source: anthropic.com ↗

Anthropic released results from an automated drug-discovery project with Citrini Research, running an internal Claude Science model that designed functional mini-protein binders against four lethal scorpion neurotoxins. Over a 64-hour run governed by a single prompt and four sequential follow-ups, Claude generated 318,484 design candidates.

From that initial generation, 211 candidate sequences were synthesized into physical proteins by Twist Bioscience and tested in wet labs by Adaptyv Bio. Seventeen designs successfully met strict binding criteria against the targeted toxins. Anthropic noted that these binders have not yet been proven to neutralize toxins in vivo, but the run proves that LLM-driven generative molecular modeling can condense protein candidate screening from months of directed evolution down to a three-day API loop.

Foursquare Equity Wiped Out in Final Acquisition

Source: teendrama.substack.com ↗

Foursquare co-founder Dennis Crowley posted an autopsy of the company’s recent sale, disclosing a complete $0 payout for both common and preferred equity. Despite scaling to more than $100 million in annual recurring revenue over its 17-year operating history, the company’s capitalization stack rendered equity compensation worthless.

The collapse stemmed from a 2019 private equity transaction used to acquire attribution provider Placed. The debt and preferred liquidation stack sat senior to all employees and founders. When growth slowed and leadership changed, the final exit valuation failed to clear the liquidation hurdle, vaporizing unexercised options, exercised stock, and early-stage common holdings.

Tuning Binary Decision Models Beats Full LLM-as-a-Judge Pipelines

Source: hamel.dev ↗

Evaluating complex agent traces using generalist models like GPT-4 or Claude 3.5 Sonnet is slow and runs up thousands of dollars in daily token burn. Analysis from Hamel Husain outlines how specialized binary decision models - including TypeSafe Jev 1.13, Perplexity Decider 1.1, and Cloudflare Clef Flash - cut latency and cost when grading whether an agent task succeeded.

The architectural catch is thresholding: almost every team consumes the raw classification probabilities using the default 0.50 decision boundary. By plotting precision-recall curves against 100 domain-specific traces, teams can calibrate thresholds to drastically minimize false negatives without paying the inference overhead of autoregressive text generation.

Synthetic Web Spam Successfully Polluted ChatGPT Citation Chains

Source: peec.ai ↗

An experiment using a synthetic programmatic site called stateglobe.com proved that generative engines ingest unverified synthetic data. The author generated 50,000 completely fabricated national metrics across 200 countries using GPT-4.1. While Google identified the low-quality generation and deindexed the domain completely, conversational AI scrapers did not.

ChatGPT surfaced the synthetic metrics directly in user search answers, driving 10,000 sessions and generating citations across 330 unique domains, including peer-reviewed academic literature and financial research notes. For machine learning engineers building Retrieval-Augmented Generation (RAG) pipelines, the case confirms that relying on public search-index retrieval without strict whitelist provenance guarantees source poisoning.

Hardware Scaling, Agent Limits, and Quick Links

Source: rameznaam.com ↗

• Sub-agent limits: Ramez Naam evaluated the theory of recursive self-improvement through horizontal agent swarms, demonstrating that parallel agent scaling hits diminishing returns on novel scientific reasoning due to compounding propagation errors.
• Orchestration evolution: Ethan Mollick tracked model capability shifts in his retrospective 30 Months on the Exponential, noting that 180 public AI models now surpass original GPT-4 benchmarks, with 19 running directly on consumer smartphones.
• Teleoperated home cleaning: A hands-on review evaluated hiring two humanoid robots from Tau Robotics for $60 an hour in San Francisco. Remote teleoperators handled basic surfaces but dropped objects, ran into battery ceilings, and proved slower than human cleaners ($44/hr).
• Career depth: Stripe engineer Vikram Singh published an essay on how junior engineers can accelerate their promotion velocity by picking an isolated, hyper-specific component (such as dashboard rendering charts) and building complete ownership over it within four months.
• Long-term pedagogy: Steve Krouse applied the Lindy Effect to programming education in What Shall I Teach My Son?, arguing why fundamental social schooling structures routinely survive technological shifts.
• Engineering notebook: A public 631-page collection of technical observations and mental models was compiled in Bits and bobs.