Get tomorrow's brief in your inbox
Friday, May 23, 2026
Today: Anthropic dominated coding conversations this week with developers shipping Claude-written code to production without reading it. Google's AI search broke badly enough that you can't search for the word "disregard." Memory chip workers are getting $340,000 bonuses from Samsung because AI data centers are eating up all the RAM production capacity.
Code with Claude: Half of developers are shipping AI-written code they haven't read
At Anthropic's London developer event this week, almost half the room admitted they've shipped code written entirely by Claude without reading it first. This is not a fringe practice anymore. It's becoming standard workflow.
Tools like Claude Code and OpenAI's Codex are pushing developers past autocomplete into full task delegation. You describe what you need, the AI writes it, runs tests, and prepares it for review. For enterprises, this is speed with control, but the "control" part is getting looser as confidence builds. Virgin Atlantic used Codex to refactor legacy codebases in 30 minutes instead of two weeks and launched their mobile app with zero P1 defects. Cisco built most of their AI Defense security platform with Codex, cutting delivery time from quarters to weeks.
The question isn't whether AI can write quality code anymore. It's whether your team is comfortable deploying agentic systems at scale when you're trusting code you didn't write or fully review. If you're still having developers write everything by hand, you're competing against teams moving 10x faster. If you're letting AI ship without human oversight, you're one bad hallucination away from a production disaster.
MIT Technology Review | OpenAI Gartner Report | Virgin Atlantic Case Study
Workday launches AI solopreneur accelerator with $10k grants
Workday, Anthropic, and LISC are funding 15 aspiring solopreneurs with $10k each, plus free Claude credits and AI training. Applications open now for a July cohort. If you're thinking about starting a business and want structured AI training plus cash to cover expenses, this is a legitimate opportunity. The catch: you'll need to demonstrate a viable business idea and commit to the curriculum and coaching.
Over 75% of small business owners already use AI tools. This program aims to help solopreneurs who are resource-constrained but determined. One-on-one coaching from LISC's Business Development Organizations is included. Small Business Trends
Nvidia's Nemotron-Labs Diffusion: faster text generation without token-by-token limits
Nvidia released diffusion language models at 3B, 8B, and 14B scales that generate multiple tokens in parallel instead of one at a time. This matters because traditional models spend most of their time loading weights from memory, not computing. Diffusion models can better leverage modern GPUs and run faster at low batch sizes.
The models support three modes: autoregressive (standard), diffusion (block-by-block generation), and self-speculation (draft with diffusion, verify with autoregressive). You can switch between modes at deployment without changing your application code. If you're building latency-sensitive apps or running single-query workloads, this is worth testing. Commercially friendly license. Hugging Face
Google's AI glasses are almost ready (but not quite)
Google showed off prototype AI glasses with in-lens displays at I/O this week. They pair with iOS and Android, show widgets for weather, directions, translations, and custom AI-generated widgets. You activate Gemini with a two-second press, it listens and can take photos.
The demo units worked, but the venue Wi-Fi was slow and it took 45 seconds to turn a photo into an anime character. Audio quality wasn't tested properly because of noise. These are not shipping soon. The audio-only version comes this fall, the display version is still in development. If you were hoping for AR glasses you could buy today, keep waiting. TechCrunch
Small specialized models are beating frontier APIs in real enterprise tasks
A 3-billion-parameter model fine-tuned for structured OCR outperformed every commercial frontier API tested, including GPT-4, Claude, and Gemini, at 50x lower cost. This result challenges the assumption that bigger is always better.
When a model's training is specialized enough for the deployment task, parameter count stops being the deciding factor. Dharma's DharmaOCR benchmark showed that a well-resourced enterprise can replicate this through fine-tuning pipelines. The quality gap ran opposite to the cost gap: the highest-scoring model was also the cheapest to operate.
If you're defaulting to the largest frontier model for every task, you're likely overpaying. The right move is to test whether a smaller specialized model can handle your specific workload. Hugging Face
Virtual accountants: cloud-based financial help without full-time hire costs
Virtual accountants are remote financial professionals who handle bookkeeping, tax prep, and financial analysis using cloud-based software. They're cost-effective compared to in-house accountants and offer flexible scaling as your business grows.
They maintain data security, help with compliance, and stay current on regulatory changes. Communication happens via video, email, and secure messaging. If you're a small business that needs financial help but can't afford or justify a full-time hire, this is a proven model. Real-time collaboration through cloud tools means you get access to your financial data from anywhere. Small Business Trends
AI memory shortage is making consumer electronics more expensive
Memory manufacturers (there are only three big ones left) have a fixed wafer capacity. Until recently, high-bandwidth memory (HBM) for GPUs got 2% of that capacity. AI data center growth has pushed it to 20% by end of 2026. A single gigabyte of HBM uses 3x the wafer capacity of regular DDR or LPDDR.
This is already hitting the sub-$100 smartphone market, which matters most in Africa and South Asia. Consumer products that use memory are getting more expensive and will stay that way for several years. Memory companies learned from rivals going extinct that you should always under-provision capacity, so don't expect them to build more fabs quickly. If you're budgeting for tech purchases, plan for price increases. Simon Willison
Samsung memory chip workers negotiated $340k bonuses
Samsung reached a tentative deal with semiconductor employees who threatened to strike. Memory chip workers on $50k salaries could get total bonuses of $416k this year. The deal includes 50% of annual salary as cash, plus 10.5% of operating profits as stock-based bonuses.
This is still smaller than SK Hynix bonuses, and Samsung made it conditional on hitting profit milestones. Samsung's earnings increased eightfold, largely from memory chip sales driven by AI demand. The union vote is pending but expected to pass. The Verge
GitHub breached by software supply chain attack group TeamPCP
GitHub confirmed a breach after a developer installed a poisoned VSCode extension. The hacker group TeamPCP claims to have accessed around 4,000 GitHub code repositories. GitHub found at least 3,800 compromised repos, all containing GitHub's own code, not customer code.
TeamPCP has carried out 20 waves of supply chain attacks in recent months, hiding malware in over 500 distinct pieces of software. They're now selling GitHub's source code on BreachForums. This is the longest-running spree of software supply chain attacks ever. If you use VSCode extensions or pull from open source repos, audit what you've installed and verify sources. Ars Technica
Google's AI search broke so badly you can't search "disregard"
Google's AI Overviews are misinterpreting action-related queries like "disregard," "ignore," and "skip." When you search "disregard," the AI Overview responds with "Got it. If you need anything else or have a new question later, just let me know!" instead of showing you the definition.
Google removed AI Overviews for "disregard" after the issue went viral. They're working on a fix. This is what happens when you foreground AI summaries and push traditional search results down the page without testing edge cases. For the first time in 15 years, Bing delivered a better search result than Google. The Verge | TechCrunch
AI startups are inflating revenue numbers with "contracted ARR"
Multiple founders and investors confirmed that AI startups are reporting contracted ARR (CARR) as regular ARR. CARR counts revenue from signed customers who aren't onboarded yet. One VC saw companies where CARR is 70% higher than actual ARR, with much of that revenue never materializing.
At least one high-profile enterprise startup reported over $100 million in ARR when only a fraction came from paying customers. The rest was from contracts not yet deployed. This matters because investors, journalists, and competitors are making decisions based on inflated numbers. If you're evaluating AI vendors, ask for actual collected revenue, not contracted revenue. TechCrunch
Grok is failing in government and enterprise adoption
Reuters reviewed federal records showing Grok or xAI appeared in only three examples of government AI use, compared to 230+ for OpenAI and dozens for Google and Anthropic. When Grok did appear, it was for basic admin work like document drafting.
Pentagon sources said staffers prefer Gemini or Claude because Grok "just isn't the best model out there." Public leaderboards confirm this. Musk recently admitted xAI used OpenAI's models to help train Grok through distillation, which is controversial when using a rival's system. SpaceX's IPO filing put AI at the heart of its pitch, claiming a $28.5 trillion addressable market, but if the product isn't competitive, those projections don't hold. The Verge
Best AI video tools by category
If you need to create video content, here's what actually works in 2026: - Generate from prompts: Google Veo for reliable results, Runway for film-making - Edit existing footage: Descript for script-based editing, OpusClip for extracting viral clips from long videos - Social media from prompts: invideo AI, revid.ai for templates - Digital avatars: Synthesia for high-quality, HeyGen for interactive
All export at least 720p, most do 4K. Focus on tools that match your workflow (editing vs generating) rather than trying to use one tool for everything. Zapier
Best AI chatbots: pick by use case, not hype
The best model for one task often falls flat for another. Don't subscribe to everything. Pick 2-3 based on what you actually do daily. Zapier
The coding world is splitting into two camps: teams shipping AI-written code at 10x speed and teams still writing everything by hand. The gap will widen fast. Meanwhile, memory shortages driven by AI data centers are pushing consumer electronics prices up across the board, and software supply chain attacks are escalating from rare events to weekly occurrences. If you're not auditing your dependencies and testing specialized models for cost savings, you're leaving money and security on the table.