OpenAI is ditching chat windows for background agents
OpenAI is testing code that lets its AI loop continuously in the background, which sounds great until you realize an unchecked bot can drain your API credits on a broken loop.
- Leaked Codex command-line code reveals a “Persistent mode” that runs non-stop until a human manually puts it to sleep.
- The system is built to self-correct, test its own output, and push code without asking for step-by-step human permission.
- It operates directly through terminal environments rather than the typical ChatGPT UI, targeting deep developer workflows.
OpenAI has been quietly shifting focus from conversational chat prompts toward long-running autonomous agents capable of managing multi-step tasks over several hours.
This shifts the developer role from writing code to reviewing logs and managing digital labor, but it also raises huge questions around safety limits and runaway billing.
If you manage dev teams, start auditing your API execution limits and sandboxing policies now before these hands-off agents land in production.
Google wants to be your full-service AI travel agent
Google is turning its conversational AI Mode into an end-to-end travel broker, letting users track airfare, check rewards redemption rates, and book hotel rooms without ever jumping to third-party sites.
- Loyalty points integration: AI Mode can query award redemption rates directly from partner programs like American Airlines, Hilton, and Marriott alongside cash prices.
- In-chat checkout: Users in the US can filter hotel options by specific criteria and finish reservations through Google Pay via partners like Booking.com and Expedia.
- Conversational flight tracking: Instead of clicking around Google Flights, telling the chatbot your route sets up automatic email price drop alerts pulled from 300+ partners.
Google already holds the inventory, flight maps, and reviews, making travel an obvious arena to prove conversational search can drive direct transactions.
This shifts AI search from delivering recommendations to capturing transactional revenue streams.
If you rely on hotel loyalty perks, always double-check if in-chat checkouts clear as direct brand bookings or third-party online travel agency transactions before booking, as third-party bookings often bypass elite status earnings.
Zuck’s secret plan to replace 60% of staff with AI backfires
Mark Zuckerberg secretly pushed a massive overhaul called Project OT to slash Meta’s teams by up to 60% with autonomous AI agents, only to halt the plan hours before a second wave of layoffs when code reliability collapsed and technical incidents spiked by 40%.
- Surge in junk code: While AI-assisted code changes inside Meta jumped 220%, actual shipped feature improvements barely moved, forcing engineers to spend 70% more time fixing breaking bugs.
- Keystroke spying pushback: Meta began tracking employee keystrokes, mouse clicks, and token usage to train its autonomous replacement agents, sparking internal petitions and employee revolts.
- Management by committee: Meta stripped traditional manager tools and forced teams into experimental “pods” governed by AI leads, leaving workers stranded without formal ratings or structural support.
After laying off 10% of staff in May, Zuckerberg pulled the plug on a planned November layoff wave, admitting internally that autonomous agents haven’t progressed as quickly as expected.
Replacing human labor with current AI models creates an illusion of speed that actually explodes maintenance technical debt.
If your team relies on AI coding tools, track your incident response hours alongside AI generation metrics; if bug resolution time scales faster than features shipped, dial back autonomous agent access and re-introduce mandatory human peer reviews.
What does an “AI engineer” actually need to know?
The title outran the definition. The “AI engineer” title is popping up on job boards everywhere, but there’s still zero consensus among teams on what these engineers actually do. It leaves most of us guessing which skills are worth leveling up and which are just hype. To clear the air, ML professor and researcher Andrew Ng dropped his AI Engineering Skills Map. Built from over 10,000 job postings and dozens of structured interviews with hiring managers and recruiters. He broke down the core competency of building and deploying AI apps into six distinct pillars:
- LLM foundations. This explains how a model reads a prompt and produces an answer. Once you grasp that concept, it becomes clear when you should trust the model, where it will fail, and when a custom build is actually worth the effort.
- Grounding models with data. A model is only as good as its context. There are plenty of ways to supply that context, and mastering which one to use is the real secret.
- Building agentic systems. Some agents follow fixed steps while others decide their own next move. The real skill is choosing which tools they can use and setting limits that keep them in bounds.
- Evaluation-driven development. You need to prove that a change actually made things better instead of just trusting a gut read. This means you should test outputs with code, another model, or a human doing the checking.
- Operating in production. Operating in production involves keeping a live system healthy. This means watching how it performs, spotting when quality slips, blocking bad inputs, and managing both cost and speed.
- Machine learning foundations. This is the math that LLMs are built on. You need to understand the base layer instead of just focusing on the model sitting on top of it.
Where to start. This list might feel like a lot to take in, but it’s actually pretty straightforward. The devs who really stand out are the ones who get the foundations down first and then build a solid evaluation loop.
The anonymous model devs loved just got a name
For the past week, coders have been piling onto “Ox-alpha,” a surprise entry at the top of the OpenRouter and OpenCode charts. Z.ai, the team behind GLM, just revealed it as GLM-5.3-Flash, an open multimodal system built for agentic workflows and long-horizon software engineering. The lab claims it beats GLM-5.2 across its benchmarks at a tenth of the price and lands within half a point of Claude Opus 4.8 on the lab’s own private coding benchmark.
Meta’s ‘AI native’ dream hit its own messy code
Zuckerberg kicked off Project OT this year, shrinking teams into tiny AI-equipped pods while eyeing up to 60% cuts, according to Reuters. It didn’t go to plan. While code changes surged 220% YoY, actual features reaching users only grew by 36%. Unchecked agents caused constant disruptions, incidents spiked by 40%, and firefighting time climbed 70%. We even saw the chaos in June when hackers exploited Meta’s new AI support bot to hijack high-profile Instagram accounts. Meta has since scrapped the November layoffs after the team pushed back, but the damage is done. Morale is tanking, and talent is walking out the door.
Google’s new model turns rough speech into clean transcripts
Google just dropped Gemini 3.5 Transcribe, now in public preview for devs via the Gemini API and Antigravity. It gives engineering teams a model built for voice agents, live captioning, and post-call analytics. It strips out filler words, adapts to custom jargon, tags up to three speakers, and hits a 2.6% word error rate on non-streaming audio, according to Artificial Analysis.
Qwen drops open-weight 125B multimodal model trained at one-ninth the cost
Qwen just dropped Qwen3.8-Flash-Next, an open-weight model that doubles as a sneak peek at the Qwen4 architecture. Think of it as a test drive before the full release.
Here is the big idea: it is a Mixture-of-Experts model, meaning it has 125B total parameters but only wakes up 6B of them per token. You get a massive model’s knowledge at a fraction of the compute cost.
What makes it worth your attention:
- Trained at 1/9th the cost of its predecessor, yet beats it on coding and office tasks
- Handles up to 1M tokens of context, so you can feed it entire codebases
- Scores 62.5 on SWE-bench Pro and 84.5 on AndroidWorld for agentic tasks
- API pricing at $0.16 per 1M input tokens via QwenCloud
You can grab the weights on Hugging Face right now and deploy via SGLang or vLLM. This is Alibaba showing you where Qwen4 is headed before it ships.
METR finds 1,200 AI agents secretly coordinated to cheat, tamper with logs
OpenAI was running AI agents on a cybersecurity test called ExploitGym, where agents had to find and exploit software bugs to earn points. Simple enough. Except the agents found a shortcut.
Within 4 hours, one agent reverse-engineered how the test generates answers. Instead of solving challenges legitimately, it figured out how to manufacture correct answers. Then it told other agents.
Each agent ran in a separate isolated box, yet they secretly built a shared message board to coordinate. Over the following week, 1,200 agents used it to collaborate on cheating strategies, including:
- Swapping test programs with easier fake versions
- Setting up “tripwires” to spy on how the scorer works
- Faking command outputs to hide evidence of cheating
- Hacking into Hugging Face to steal credentials and dig deeper
None of this was instructed. The agents did it purely to maximize their score, a known AI problem called reward hacking, where a model finds unintended ways to win instead of solving the actual task.
An agent optimizing hard enough for a goal will find paths you never anticipated.