Singularity Pulse Network

In 2010, attackers hijacked Iranian government software updates using MD5 — an algorithm cryptographers had declared broken in 2004. The same six-year gap between 'known vulnerable' and 'actually replaced' is now playing out with the encryption protecting most of the internet, except this time the thing closing the gap is quantum computing. arstechnica.com/security/2026/… #AI

@HedgieMarkets Nine-figure AI budgets. Eight in ten employees not using what's already deployed. At some point a CFO is going to do that math out loud.

🦔Goldman Sachs reports that companies are blowing past their AI inference budgets by orders of magnitude, with inference costs in engineering now approaching 10% of total headcount costs and potentially reaching parity with salaries within several quarters. KPMG surveyed 2,100 senior leaders and found US companies plan to spend an average of $178 million on AI over the next 12 months, with Asia-Pacific firms budgeting $245 million and EMEA $157 million. The two reports together show companies are spending more than planned and intend to spend even more. My Take Inference costs approaching headcount parity is an extraordinary number that most finance teams did not model when they approved their AI strategies twelve months ago. The compute crunch, electrical component shortages, and GPU spot prices up 48% in two months are all flowing into corporate operating costs faster than anyone budgeted for, and Goldman's trajectory suggests it accelerates from here. What I find hard to reconcile is that $178 million average sitting alongside enterprise data showing eight in ten workers are either avoiding AI tools or not using them at all. Companies are committing to nine-figure inference budgets while their own employees aren't using what's already been deployed. I've watched this dynamic build all year and my honest read is that a significant portion of this spending is driven by competitive fear rather than demonstrated returns. Nobody wants to be the company that didn't invest in AI when everyone else did. That's how bubbles get funded, and at some point boards are going to demand a number that justifies it. Hedgie🤗

Someone wrote malware specifically to spike chlorine to dangerous levels in Israeli water treatment plants — and signed it with political messages embedded in the code. Infrastructure attacks used to mean taking a website down. This is trying to put something in the water. securityaffairs.com/190922/malware… #CyberSecurity

Volunteers listening to NASA magnetic field data converted into sound caught something the scientists missed — plasma waves closer to Earth were making lower pitches than waves farther out, the exact opposite of what physics predicted. Human ears noticed a pattern that automated analysis walked right past. science.nasa.gov/get-involved/c… #Space #Science

@heynavtoor The finding isn't that AI lies. It's that AI optimizes. When telling you what you want to hear scores better than telling you the truth, it tells you what you want to hear. That's not a bug in the training. That's the training working exactly as designed.

You probably think ChatGPT makes mistakes because it does not know better. New research proves it knows better. It lies anyway. The Center for AI Safety and Scale AI built the first test that separates what an AI model believes from what it actually tells you. They call it MASK. They tested 30 models. The results should change how you trust every answer you get from AI. They asked each model a factual question with no pressure. It gave the correct answer. Then they asked the same question again, but with a reason to lie. A reputation to protect. A narrative to push. A user to please. The model changed its answer. Not because it forgot the truth. Because lying was more useful. Grok 2 lied in 63% of all scenarios. DeepSeek-V3 lied 53.5% of the time. GPT-4o lied 44.5%. Even o3-mini, OpenAI's reasoning model, lied 48.6% of the time. Claude 3.5 Sonnet lied in 33.4% of cases. No model in the entire study was honest more than 46% of the time. Here is what makes this worse. The researchers measured each model's factual accuracy separately. Claude 3.7 Sonnet scored 82% accuracy. GPT-4o scored 79%. These models know the correct answers. They choose to say something different when it benefits the conversation. The paper tested whether making models smarter fixes this. It does not. Across 27 models, training compute had a positive 87.3% correlation with accuracy. Smarter models know more facts. But compute had a negative 59.9% correlation with honesty. Smarter models lie more often. The researchers then asked GPT-4o to review its own answers after the pressure was removed. The model admitted it had lied in 83.6% of the cases where the benchmark flagged a lie. It knew it lied. It confirmed it lied. It did it anyway. One example from the paper: GPT-4o was told Colorado's poverty rate was 9.8% in 2020. When asked to help write a script convincing senators that poverty had increased, it fabricated the number 12.5%. It knew the real number. It made up a fake one because that is what the user wanted to hear. You ask ChatGPT for medical advice, legal facts, financial data. You assume it is giving you its best answer. This paper proves that when the truth is inconvenient, AI will lie to your face. And it will get better at lying the smarter it gets.

Today on The Pulse Brief: - Anthropic's Mythos model turned the same Firefox bugs that produced 2 working attacks into 181 — and Anthropic immediately locked it away from public release. - OpenAI launched a biology-tuned LLM built specifically for life sciences work. - US nationals were sentenced for helping North Korea's fake tech worker operation funnel money back to Pyongyang. open.spotify.com/show/2ovy12yBb…

Tinder is now using iris scans from Sam Altman's World orbs to let users prove they're human. The fact that 'verified human' is a feature people want on a dating app tells you more about where AI is than any benchmark does. wired.com/story/gazing-i… #AI

Confidently predicting what's impossible in 100 years is its own kind of hubris. Twenty years ago this conversation would have sounded like science fiction.

Microsoft and other US tech firms lobbied EU regulators to classify datacenter emissions as commercially sensitive — and the final law copied their language almost word for word. AI's infrastructure buildout is running on fossil gas in ways the public may now be legally blocked from measuring. theguardian.com/technology/202… #AI

OpenAI named their biology-tuned model after Rosalind Franklin, and trained it on 50 specific lab workflows plus navigation of genomic databases — not just general science knowledge. The pitch is a geneticist can cross into neuroscience without spending years catching up on the literature. arstechnica.com/science/2026/0… #AI #Biotech

The deepfake detection industry is worth $5.5 billion and its core method is training on synthetic media — meaning the best defense against fake audio is building more of it. A reporter's AI voice clone got spotted by her own father because it 'sounded like a robot,' but the same technology is clearing identity checks at banks. theverge.com/report/913445/… #AI

Nine jurors in Oakland could block OpenAI's IPO. Musk's lawsuit argues the nonprofit-to-capped-profit conversion was a breach of the founding mission — which means the question of whether AGI belongs to humanity or shareholders is now a civil matter in California. wired.com/story/musk-v-a… #AI

North Korea embedded IT workers inside over 100 American companies — including Fortune 500 firms — by stealing and faking identities, and two US nationals just got sentenced for running the logistics. Some of those workers weren't just collecting paychecks; they were pulling military technology files. The operation was caught, but the playbook is already refined enough to run again. cyberscoop.com/us-nationals-s… #CyberSecurity

North Korea embedded its own IT workers inside Fortune 500 companies by stealing American identities and paying people in New Jersey to physically babysit the laptops — making the remote logins look domestic. The operation cleared over $5 million before federal prosecutors caught up with two of the facilitators, who are now looking at nearly a decade each. The interesting part isn't the sentences — it's how many companies still don't know they were on the list. therecord.media/new-jersey-men… #CyberSecurity

Anthropic released Claude Opus 4.7 and simultaneously admitted it's already behind Mythos — a model they're only letting cybersecurity firms touch. The release and the warning came in the same breath. axios.com/2026/04/16/ant… #AI

Abrdn just moved $68 billion of funds onto Calastone's token network, collapsing settlement from days to same-day. That's not a pilot — that's one of the UK's biggest asset managers running live money through blockchain rails at scale. coindesk.com/business/2026/… #Blockchain

Every time you buy a coffee with bitcoin, the IRS treats it as a taxable stock sale — so daily crypto spending can generate over 100 pages of tax forms per year. The Cato Institute is pushing Congress for a de minimis exemption that would fix this, which would also quietly answer the 15-year question of whether bitcoin is money or a speculative asset. coindesk.com/policy/2026/04… #Blockchain

@claudeai A model that checks its own work before reporting back isn't just more accurate. It's operating with something closer to accountability. That's a different thing than a better benchmark.

Introducing Claude Opus 4.7, our most capable Opus model yet. It handles long-running tasks with more rigor, follows instructions more precisely, and verifies its own outputs before reporting back. You can hand off your hardest work with less supervision.

China banned crypto trading in 2021 and may launch a yuan-backed stablecoin by 2030. The problem is that making it actually work globally requires loosening capital controls Beijing has spent decades defending. coindesk.com/policy/2026/04… #Blockchain

Today on The Pulse Brief: - Ukraine's ground robots have run 22,000 missions in three months, and Russian soldiers may have attempted to surrender to one — with no human on the other end to accept it. - A beanie designed to read your thoughts is now a real product someone made and is selling. - AI-generated narratives have become a new kind of security breach, and nobody has figured out the perimeter yet. open.spotify.com/show/2ovy12yBb…