SIGNAL

Wednesday, March 25, 2026

16 stories · 5 min read

THE SIGNAL

The pattern is unmistakable: frontier AI labs are abandoning consumer bets and consolidating around infrastructure and internal capability. OpenAI's pivot away from Sora signals what we've quietly known for months—generative video isn't the wedge product that unlocks AGI, scaling and reasoning are. Watch where the talent and capital actually flow; that's where the real race lives.

★ Must ReadOpenAI just gave up on Sora and its billion-dollar Disney deal

OpenAI has discontinued Sora, its video generation tool launched in late 2024, shutting down both the consumer app and developer API with no integration into ChatGPT planned. The shutdown dissolves the licensing agreement with Disney that was signed just months after Sora's launch, representing a significant retreat from a major commercial partnership. The decision signals either technical limitations in competing with rivals like Google's Veo or insufficient commercial traction to justify continued investment. This marks a rare public abandonment of a flagship product by OpenAI and raises questions about the viability of generative video as a near-term revenue driver for the company.

The Verge AI· Richard Lawler· 2 min· 59d agoSource →

01QUICK HITS

Is anybody else bored of talking about AI?

Trending on Hacker News with 576 points and 402 comments.

Hacker News · 1 min

→03

Tell HN: Litellm 1.82.7 and 1.82.8 on PyPI are compromised

About an hour ago new versions have been deployed to PyPI. I was just setting up a new project, and things behaved weirdly. My laptop ran out of RAM, it looked like a forkbomb was running.

Hacker News · 1 min

→04

Wine 11 rewrites how Linux runs Windows games at kernel with massive speed gains

Trending on Hacker News with 731 points and 256 comments.

Hacker News · 1 min

→05

Show HN: Gemini can now natively embed video, so I built sub-second video search

Gemini Embedding 2 can project raw video directly into a 768-dimensional vector space alongside text. No transcription, no frame captioning, no intermediate text. A query like "green car cutting me off" is directly comparable to a 30-second video clip at the vector level.

Hacker News · 1 min

→

03THE TENSION2▸

PERSPECTIVE A

“LaGuardia pilots raised safety alarms months before deadly runway crash”

— Hacker News

Two pilots were killed in a collision between an aircraft and ground vehicle at LaGuardia Airport, an incident preceded by documented safety concerns raised by pilots months earlier. The fatality represents a potential failure in the safety reporting and corrective action pipeline, suggesting known hazards were not adequately addressed before the incident occurred. This pattern—where pre-incident warnings go unheeded—typically indicates systemic gaps in either safety culture, resource allocation, or management responsiveness to frontline personnel. The case will likely prompt regulatory scrutiny of LaGuardia's safety protocols and incident reporting procedures.

Source →

♦vs♦

PERSPECTIVE B

“Self-propagating malware poisons open source software and wipes Iran-based machines”

— Ars Technica AI

A self-propagating malware variant has infiltrated open source software repositories and selectively targeted Iran-based systems for destruction. The malware spreads through compromised dependencies in development environments, allowing it to move laterally across networks before executing destructive payloads on machines matching specific geographic or organizational criteria. This represents a convergence of supply chain vulnerability and precision targeting—attackers can reach global development teams while discriminating victims at execution time. Organizations need immediate network audits focused on dependency integrity and execution anomalies, particularly those using affected open source packages.

Source →

04RESEARCH1▸

[AINews] Dreamer joins Meta Superintelligence Labs — 9 month retro of Personal Superintelligence

Dreamer, a researcher featured on the Latent Space podcast just 11 days prior, has been hired by Meta's Superintelligence Labs (MSL), suggesting rapid talent acquisition around emerging AI capability work. The speed of the hire—announced days after the podcast episode shipped—indicates either pre-existing recruitment discussions or strong enough demonstrated expertise to accelerate hiring timelines. This reflects broader competition among top AI labs to secure specialized talent in superintelligence research, a domain gaining significant corporate investment and focus.

Latent Space· 3 min· 60d agoSource →

05BUILDERS3▸

★ Must ReadI’m sharing our internal AI engineering cheatsheets

A Towards AI engineer has publicly released internal markdown documentation covering AI engineering practices for writing, coding, and agent development. The cheatsheets represent the specific operational templates and guidelines the organization uses internally, providing concrete patterns rather than theoretical frameworks. This move signals increasing normalization of sharing engineering standards across the industry, potentially accelerating adoption of common practices while reducing the knowledge gap between established AI teams and practitioners. For organizations building AI systems, the release offers immediately applicable reference material for standardizing prompt engineering, code patterns, and agent architecture decisions.

Louis Bouchard· Louis-François Bouchard· 2 min· 59d agoSource →

Did You Know AI Can Do… That?

The article argues that current chatbot interfaces—text-based input boxes and cursor prompts—are constraining how organizations explore AI capabilities and applications. The implication is that AI systems can perform significantly broader functions than typical conversational use cases suggest, but users aren't discovering them because the interface design defaults to chat-like interaction. This matters because interface limitations may be suppressing organizational adoption and innovation; expanding how users interact with AI (voice, visual, API-first, embedded workflows) could unlock value currently hidden behind the assumption that "AI" means "chatbot conversation."

Leon Furze· 2 min· 59d agoSource →

The Future of On-Device AI

Liquid AI has developed a compact edge AI model that achieves performance levels previously requiring much larger systems, using architectural innovations that reduce computational overhead while maintaining accuracy. The company's approach leverages structured state-space models and optimized parameter efficiency, enabling deployment on devices with limited processing power and memory. This advancement matters because it shifts the economics of AI deployment—reducing reliance on cloud infrastructure, lowering latency, and enabling real-time processing on smartphones and IoT devices, which has significant implications for privacy, cost, and the competitive landscape for model development.

AI Made Simple· Devansh· 2 min· 60d agoSource →

06GO DEEP1▸

★ Must Read🔬Why There Is No "AlphaFold for Materials" — AI for Materials Discovery with Heather Kulik

Despite AlphaFold's breakthrough in protein structure prediction, materials science hasn't achieved an equivalent AI breakthrough because materials discovery involves fundamentally different constraints: proteins fold into discrete stable structures, while materials exist across continuous property spaces with billions of possible compositions and configurations. According to Heather Kulik's analysis, the field requires not just better prediction models but domain-specific approaches—such as physics-informed ML and active learning loops—that account for synthesis feasibility, manufacturability, and performance trade-offs that don't apply to protein folding. This distinction matters because it explains why generalizable AI won't simply transfer across scientific domains, and why materials companies must build narrower, application-specific systems rather than waiting for a universal "AlphaFold moment."

Latent Space· Brandon Anderson· 3 min· 59d agolong readSource →

07VOICES3▸

“We're saying goodbye to Sora,”

— Richard Lawler, The Verge AI