Software Engineer

Amazon trims artificial general intelligence team in latest workforce cuts. By Jonathan Easton 23/07/2026 Amazon has cut jobs in parts of its artificial general intelligence (AGI) organisation as the technology giant continues restructuring its workforce while investing heavily in AI development ahead of its second-quarter earnings next week. The company did not disclose how many employees were affected or which teams bore the brunt of the reductions. In a statement issued after enquiries from Reuters and CNBC, an Amazon spokesperson said the company was "sharpening our focus on the initiatives that matter most for customers" and that this had led to "some difficult decisions, including eliminating some roles within parts of its AGI organisation", while continuing to invest in its most important AI priorities. The cuts represent the latest in a series of targeted layoffs following Amazon's broader workforce reduction of around 16,000 jobs in January. According to Reuters, employees reporting to Adeeb Shanaa, vice president of artificial general intelligence data services, and Vishal Sharma, vice president of AGI information, said on online forums that their teams had been affected, although the full extent of the job losses remains unclear. According to CNBC, some of the employees laid off worked on model customisation and post-training, while the AGI unit remains central to Amazon's AI ambitions. The division develops foundation models including the Nova family and oversees work spanning silicon development and quantum computing, as the company seeks to compete with rivals developing increasingly capable AI systems. The scale of the disruption became clearer as affected employees began posting publicly. Several members of the AGI foundation team said on X that they had been let go, offering a rare glimpse inside teams Amazon declined to detail. Yuxin Tang, an LLM pretraining researcher, wrote that "most members in our MoE pre-training team got impacted today", referring to the mixture-of-experts approach used to train large models efficiently. Miao Xiong, whose work focused on data curation for pretraining, captured the mood of the day. "My job was deciding which data points matter for pretraining. Turns out I was the one that got filtered out," she wrote, adding that a reminder for the team's weekly project sync still went off at 1pm "for a project that no longer has a team". The cuts also reached newer hires: Kyle Wong, who worked on computer-use agents, said he was laid off the day after his 23rd birthday, having joined from work on GUI agents at startup Simular and at Apple. The public posts, many of which attracted hundreds of replies and job offers from rival labs, underline how sought-after such specialists remain even as Amazon trims the teams behind its foundation models. The restructuring follows significant leadership changes within Amazon's AI business over the past year. Reuters reported that Rohit Prasad, who previously oversaw the AGI group, left the company at the end of 2025, with responsibility transferred to senior vice president Peter DeSantis as part of a broader technology organisation. David Luan, who led Amazon's AGI Lab after joining through the acquisition of AI startup Adept, departed in February. Peter DeSantis acknowledged in an interview with CNBC last month that Amazon's AI models "haven't been at the very frontier for the very largest, most demanding workloads". He added that the company was working to strengthen its technology and hoped to build one of the "most capable intelligent models out there". The layoffs come as Amazon continues an aggressive expansion of its AI infrastructure. CNBC reported the company has forecast capital expenditure of $200 billion this year, more than 50 per cent higher than in 2025, while raising tens of billions of dollars in debt to support those investments.

Control your AI assistant from WhatsApp: A practical guide. Anmol Singh Meet Dume Cowork A desktop AI agent that handles files, browser research, and reports end-to-end. Your AI assistant doesn't need a screen to be useful. If you can send a text message, you can already delegate a task - and that's the entire premise behind controlling an AI assistant from WhatsApp. Instead of opening an app, waiting for a laptop to wake up, or hunting for the right browser tab, you type a message the same way you'd text a colleague. Dume answers back in the same thread. WhatsApp reaches 3.3 billion monthly active users worldwide, and more than 2.3 billion people open it every single day (Demandsage, 2026). For most people, it's already the fastest way to reach another human. This guide walks through what WhatsApp control actually does, how to set it up, and where its limits are - no hype, just the practical version. Key Takeaways - WhatsApp has 2.3 billion daily active users, making it the fastest channel most people already have open (Demandsage, 2026). - You can delegate tasks, request morning briefings, and forward email threads for a summary - all inside a normal WhatsApp chat. - Setup takes under five minutes: link your number, verify, and start messaging your assistant directly. - Complex file edits and multi-app workflows still need the desktop app or browser - WhatsApp handles quick delegation, not deep work. - As of mid-2026, none of Dume's closest competitors - Sintra, alfred_, Simular, and Otto - offer native two-way WhatsApp control. Why WhatsApp control matters. Most AI assistants assume you're at a desk. That assumption breaks down the moment you're walking to a meeting, standing in line, or riding the train - which, it turns out, is most of the day. Seventy-six percent of people respond to a phone notification within five minutes of receiving it, and 71% of employees already feel pressure to stay reachable outside the office (SciTechToday workplace data, 2026). WhatsApp is where that reachability already lives. Building AI control into WhatsApp instead of a separate app removes a step nobody asked for. There's no login screen, no new interface to learn, no notification permission to grant. You already have the app open; the assistant just becomes another contact. That's a genuinely different design decision than a dashboard-first product, and it's why Dume treats WhatsApp as a primary interface rather than a bolt-on notification channel. It also matters competitively. [UNIQUE INSIGHT] Dume checked the four AI assistants most often compared with Dume - Sintra, alfred_, Simular, and Otto - and as of mid-2026, none of them offer two-way WhatsApp control. Sintra's own help center confirms WhatsApp isn't a supported channel (Sintra Help Center, 2026), and alfred_ documents a one-way SMS morning brief but doesn't list WhatsApp as a two-way channel. If WhatsApp is where your day already happens, that's a meaningful gap in the alternatives. What you can actually do. WhatsApp control isn't a chatbot bolted onto a landing page - it's the same assistant you'd use in the browser, just reachable somewhere else. In practice, that opens a specific set of moves that fit naturally into a WhatsApp thread: * Text a task while commuting: "Reschedule my 2pm with Jordan to Thursday" sent from the train updates your calendar without opening a laptop. * Get a morning briefing as a message: a daily WhatsApp text with your calendar, top three emails, and open tasks, waiting when you wake up. * Forward an email thread for a summary: forward a long thread and ask for the three-sentence version before you reply. * Send a voice note instead of typing: dictate a task walking between meetings and let the assistant transcribe and act on it. * Ask follow-up questions in-thread: "did that reply to Jordan actually send?" without switching context or opening an inbox. * Approve or reject a drafted reply: the assistant proposes an email, you approve with a one-word reply. None of this requires installing a new app on your phone. It's the WhatsApp thread you already have open, just pointed at your assistant instead of a person. Setting it up. Getting WhatsApp control running takes about five minutes, and none of it requires a developer. * Open your Dume account settings and select "Connect WhatsApp." * Enter the phone number you want to message your assistant from and confirm the one-time verification code. * Send a first test message - something simple like "what's on my calendar today?" confirms the connection is live. * Set your notification preferences, including whether you want a proactive morning briefing or reply-only behavior. * Add any other authorized numbers, for example a shared assistant line for a co-founder or an EA. Once connected, the assistant recognizes your number automatically. There's no separate login for the WhatsApp channel - it inherits the same permissions, calendar access, and email connections you already set up in the browser or desktop app. What WhatsApp control can't do (yet). WhatsApp control is genuinely useful, but it isn't a replacement for the full assistant experience - and pretending otherwise would undercut the point of this guide. Multi-step file work - editing a spreadsheet, restructuring a slide deck, or running a batch file rename - still needs Dume Cowork on the desktop, where the assistant has direct access to your files and screen. Long-form drafting, like writing a full report from scratch, is also easier in the browser, where you can see the whole document as it's generated instead of scrolling a chat thread. Attachments have practical limits too: WhatsApp compresses images and strips some file metadata, so anything that needs to stay pixel-perfect is safer sent through email or the desktop app. And because WhatsApp is a personal messaging app first, sensitive documents should go through the same access controls you'd apply to any other channel - WhatsApp adds convenience, not new security guarantees. Think of WhatsApp control as the fast lane for quick delegation, not the whole highway. Frequently asked questions. WhatsApp control is one interface in a broader set. Pair it with Dume Cowork for desktop file automation, or dial in directly with AI phone calls when a task needs a real conversation instead of a text thread. Wherever your day happens, that's where the assistant should be too. Meet Dume Cowork A desktop AI agent that handles files, browser research, and reports end-to-end.