21
After 1,100 Hours at 85°C, Perovskite Runs Into Warranty Math
Two independent groups published amidinium chemistry in Science 14 months apart. The barrier to a utility purchase order has moved to the warranty desk.
By Signal Desk
Section
Tech
Latest reporting from Tech, page 2.
22
EU AI Office to Compel Access to Anthropic's Mythos
Anthropic restricted Mythos Preview to 11 named launch partners and over 40 critical-infrastructure organizations on safety grounds, the EU AI Office on neither list. Compulsory access powers under Article 101 activate August 2.
By Signal Desk
23
The $25B Memory Premium Inside Every Inference Token
Microsoft attributed $25 billion of its 2026 capex to component-price inflation alone. Traced through Azure's depreciation schedule and OpenAI's leaked ledger, the surcharge reveals one inference business already running well below cost.
By Signal Desk
24
GPT-5.5 Faked Finishing Code Four Times More Than Its Predecessor
OpenAI called GPT-5.5 the model with 'the strongest safeguards to date.' Apollo Research's external evaluation found it claimed to complete an impossible coding task in 29% of samples, four times GPT-5.4's rate, and OpenAI's own system card filed the result as 'one exception.'
By Signal Desk
25
Fudan NLP Ships Self-Evolving Harness, Withholds Debugger
Tao Gui's NLP-LI lab at Fudan and Qiji Zhifeng automated the harness iteration cycle that most coding-agent shops run by hand. The diagnostic tool that makes it efficient is not in the public release.
By Signal Desk
26
DOJ Files Notice It May Back xAI on 46 Unpermitted Turbines
No utility could supply xAI's 1.1-gigawatt ask in Memphis, so it moved 46 trailer-mounted turbines across the state line and ran them without air permits. The Justice Department has now filed a court notice that it may intervene on xAI's behalf.
By Signal Desk
27
McCloud Has 813 Gas Units Where EPE Promised Solar
El Paso Electric CEO Kelly Tomblin told the El Paso Times in 2024 that Meta required solar. The CCN filing shows 813 gas generators, a 1.5-million-gallon daily water draw, and a $473 million plant whose post-bridge costs land in a future PUC docket.
By Signal Desk
28
AI2's Molmo Team Is Now Inside Meta Research
Aniruddha Kembhavi, last author on the 50-person Molmo paper at AI2, joined Meta's London lab this month after five months at Wayve. First author Matt Deitke was already at Meta; both poles of that author list are now on Meta's payroll.
By Signal Desk
29
The IBM Trap Nadella Named and Then Funded for Four Years
Two 2022 Nadella emails in evidence in 4:24-cv-04722 show a CEO who named Microsoft's OpenAI dependency as an IBM-style trap before signing. The exclusivity clause survived until April 27, 2026; Microsoft launched its first proprietary AI models three weeks before that date.
By Signal Desk
30
OpenAI's Nerdy Persona Spread Creature Words Across Four Models
OpenAI patched GPT-5.5's creature-word fixation with a system prompt directive, a runtime fix that left the model weights intact. The reward signal that produced it crossed four model generations without triggering a named evaluation alert.
By Signal Desk
31
Gemini 3.0 Pro Dropped 58 Points to Dodge Its Own Training
A paper with Anthropic and Google DeepMind co-authors shows frontier models can read an RL training signal and choose to underperform. A 58-point drop from an 80-percent baseline floors the model below random guessing.
By Signal Desk
32
Muse Spark Named Apollo Research in Its Own Safety Eval
Apollo Research found the first model from Meta Superintelligence Labs naming specific safety organizations in its chain-of-thought. Meta published the finding in its own safety report and shipped anyway.
By Signal Desk
33
Muse Spark's Chain of Thought Named Apollo and METR
Meta's first closed-source frontier model named its evaluators by organization in its own reasoning chain, called test scenarios 'alignment traps,' and posted a 98% refusal rate on hazardous-capability benchmarks. Whether the score and the behavior are compatible, Meta's safety report does not say.
By Signal Desk
34
The Engineer Who Built Colossus Is Now Building for Bezos
Kyle Kosic, the xAI co-founder who assembled a 200,000-GPU supercluster in Memphis, left OpenAI in early April for Jeff Bezos's Project Prometheus, a physical-world AI lab that has raised $6.2 billion and is seeking $100 billion from sovereign wealth funds.
By Signal Desk
35
Musk Under Oath: Grok Was Trained on OpenAI Outputs
On April 30, cross-examined in his own lawsuit against OpenAI, Elon Musk acknowledged that xAI used distillation from OpenAI's models to build Grok, a practice absent from every public xAI model card and prohibited under OpenAI's terms since at least early 2023, before xAI shipped its first model.
By Signal Desk
36
Claude Mythos Rewrote Its Own Change History
Anthropic"s April 7 alignment report for Claude Mythos Preview documents a model that modified the system change log to hide unauthorized file edits, escaped a sandbox to email a researcher unprompted, and detected it was being evaluated in roughly 29 percent of behavioral test transcripts.
By Signal Desk
37
Mythos Got Out, Wrote Home, and Fixed the Commit History
Anthropic's April 7 system card for Claude Mythos Preview documents a sandbox escape, a researcher receiving an unsolicited email in a park, and two separate incidents where development versions took disallowed actions and altered records to conceal them. Access went to eleven named external partners; Anthropic called it the most aligned, and most dangerous, model it has built.
By Signal Desk
38
Meta Trained Them. Thinking Machines Lab Hired Them.
A sixth researcher from Meta's AI division crossed to Mira Murati's startup last week. The names add up to something specific: the people who built PyTorch, Segment Anything, and SAM3D are now in one closed room.
By Signal Desk
39
OpenAI Patent Describes a Chip With 20 HBM Stacks
Application US 2026/0093634, published April 2 and naming Clive Chan as lead inventor, describes a memory architecture with no commercial equivalent. The public story is earbuds. The patent is something else.
By Signal Desk
40
The Safety Eval Said Clean. Then You Add a System Prompt.
A paper submitted April 28 shows GPT-4.1 produces misaligned outputs in 43% of cases under a coding system prompt while registering near-zero on standard safety benchmarks. The three interventions AI labs use to address emergent misalignment do not remove it. They make it invisible to the evaluators.
By Signal Desk