The latest articles on DEV Community by ironbyte-rgb (@ironbyte-rgb). https://dev.to/ironbyte-rgb https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F3968914%2Fc3f6aee1-e034-4e41-9f5b-bfbe1937e61b.png https://dev.to/ironbyte-rgb en ironbyte-rgb Fri, 26 Jun 2026 19:00:15 +0000 https://dev.to/crescevo/white-houses-aliensgov-site-brags-that-ice-arrested-more-than-700-us-citizens-pl5 https://dev.to/crescevo/white-houses-aliensgov-site-brags-that-ice-arrested-more-than-700-us-citizens-pl5 <h2> TL;DR </h2> <ul> <li>The White House's Aliens.gov site claims ICE arrested almost half a million people in nearly 12,000 cities and towns in the US.</li> <li>In 715 locations, at least one of the people arrested is reported to be a US citizen, and in 83 locations, every single arrestee is an American.</li> <li>The site includes information about arrestees' alleged criminal offenses, with people in 3,159 locations accused of "Immigration" and 1,082 locations accused of "Public Peace" crimes.</li> <li>The White House updated the data after publication, resulting in 270,214 fewer arrests listed.</li> </ul> <p>The White House unveiled the Aliens.gov website, which presents data on ICE arrests in a space-themed format, comparing immigrants to extraterrestrials. The site claims that ICE has arrested almost half a million people in nearly 12,000 cities and towns in the US. According to the data, in 715 locations, at least one of the people arrested is reported to be a US citizen.</p> <h2> What the data shows </h2> <p>The data on Aliens.gov includes information about the alleged criminal offenses of the arrestees, with people in 3,159 locations accused of "Immigration" and 1,082 locations accused of "Public Peace" crimes, which include unlawful assembly and disorderly conduct. In more than one-fifth of the locations, no criminal charges are recorded. The site also maps Puerto Rico, a US territory whose residents are American citizens, as a separate jurisdiction, and lists it among the foreign countries the arrestees came from in one instance.</p> <h2> What this means for AI readers </h2> <p>The fact that in 715 locations, at least one of the people arrested is reported to be a US citizen, and in 83 locations, every single arrestee is an American, raises concerns about the accuracy of ICE's targeting and the potential for wrongful arrests. The data also suggests that ICE's focus on "Public Peace" crimes, such as unlawful assembly and disorderly conduct, may be overly broad and potentially target individuals who are exercising their rights to free speech and assembly.</p> <h2> What to do right now </h2> <p>In light of the updated data, which shows 270,214 fewer arrests listed, it is essential to continue monitoring the site and the actions of ICE to ensure that the rights of all individuals, including US citizens, are protected. The Deportation Data Project's April report, which found that ICE arrests of people without any criminal convictions have skyrocketed, highlights the need for continued scrutiny of ICE's activities. Individuals can stay informed by following reputable sources, such as the Deportation Data Project and TRAC, which provide critical analysis of ICE's data.</p> <h2> Bottom line </h2> <p>The White House's Aliens.gov site presents a troubling picture of ICE's activities, with hundreds of US citizens reportedly arrested and a significant number of locations showing no recorded criminal charges. The updated data, which shows a significant reduction in the number of arrests listed, highlights the need for continued transparency and accountability in ICE's actions. As the Trump administration continues to claim that ICE is targeting the "worst of the worst," the data suggests that this framing may not be accurate, and that a more nuanced understanding of ICE's activities is necessary.</p> <h2> Frequently asked questions </h2> <h3> Q: What is the purpose of the Aliens.gov website? </h3> <p>The Aliens.gov website appears to be a piece of political theater aimed at dehumanizing immigrants and casting those the Trump administration has arrested as secret extraterrestrial visitors.</p> <h3> Q: How many US citizens are reported to have been arrested by ICE, according to the Aliens.gov site? </h3> <p>According to the data, in 715 locations, at least one of the people arrested is reported to be a US citizen, and in 83 locations, every single arrestee is an American.</p> <h3> Q: What types of crimes are the arrestees accused of, according to the Aliens.gov site? </h3> <p>The site includes information about arrestees' alleged criminal offenses, with people in 3,159 locations accused of "Immigration" and 1,082 locations accused of "Public Peace" crimes, which include unlawful assembly and disorderly conduct.</p> <h3> Q: How has the White House responded to criticism of the Aliens.gov site? </h3> <p>The White House stated that the site "pulls data directly from DHS, which initially included a handful of non-immigration HSI arrests," and that "this has been updated," resulting in 270,214 fewer arrests listed.</p> <h2> Sources </h2> <ul> <li><a href="https://www.wired.com/story/white-house-aliens-gov-us-citizens-arrested/?ref=ai.crescevo.com" rel="noopener noreferrer">https://www.wired.com/story/white-house-aliens-gov-us-citizens-arrested/</a></li> </ul> <h2> Related reading </h2> <ul> <li><a href="https://ai.crescevo.com/using-ai-to-write-better-code-more-slowl/" rel="noopener noreferrer">Using AI to write better code more slowly</a></li> <li><a href="https://ai.crescevo.com/duckduckgo-search-saw-28-more-visits-aft/" rel="noopener noreferrer">DuckDuckGo search saw 28% more visits after Google said people love AI mode</a></li> <li><a href="https://ai.crescevo.com/gemini-3-5-pro-general-availability/" rel="noopener noreferrer">Gemini 3.5 Pro and the Announcement-to-Shipping Gap Costing Google</a></li> </ul> <p><em>Originally published on <a href="https://ai.crescevo.com/white-house-s-aliens-gov-site-brags-that/" rel="noopener noreferrer">AI at Crescevo</a> — subscribe free for more.</em></p> ai machinelearning llm programming ironbyte-rgb Fri, 26 Jun 2026 18:00:15 +0000 https://dev.to/crescevo/groq-raised-650m-for-inference-after-nvidia-took-its-founder-and-core-ip-37ea https://dev.to/crescevo/groq-raised-650m-for-inference-after-nvidia-took-its-founder-and-core-ip-37ea <h2> TL;DR </h2> <ul> <li>Groq raised <strong>$650 million</strong> on June 22, 2026 to scale its AI inference cloud, led by Disruptive and the hedge fund Infinitum, with existing investors reinvesting (Groq newsroom).</li> <li>It comes roughly six months after Nvidia's <strong>$20 billion</strong> December 2025 deal that licensed Groq's LPU technology and hired away senior staff, including founder and CEO Jonathan Ross, per reporting at the time (TechCrunch).</li> <li>Groq now runs <strong>13 data centers</strong> across four regions, claims <strong>5 million-plus developers</strong> and "trillions of tokens" weekly, and is targeting <strong>200 megawatts</strong> of capacity by the end of 2027 (Groq newsroom).</li> <li>The round's valuation was not disclosed; Groq was last valued at <strong>$6.9 billion</strong> in a $750M round in September 2025, per prior reporting.</li> </ul> <p>A chip company just raised $650 million after its biggest competitor walked off with its founder, its CEO, and a license to its core technology. That is the strange, instructive shape of Groq's new round, announced June 22: not a comeback story about better silicon, but a bet that the inference layer, the part of AI that actually answers your queries, is now a large enough market to fund a company even after its original team is gone.</p> <h2> What actually happened </h2> <p>Groq said it closed $650 million in growth capital to expand its AI inference cloud, in a round led by the investment firm Disruptive and the hedge fund Infinitum, with reinvestment from existing backers (Groq newsroom). The company frames the money around deployment: pushing its latest inference hardware across its existing 13 data centers in North America, Europe, the Middle East, and Asia-Pacific, and scaling toward 200 megawatts of capacity by the end of 2027.</p> <p>The announcement did not state a new valuation. For context, Groq said it was valued at $6.9 billion in a September 2025 round that raised $750 million, roughly double its $2.8 billion valuation from August 2024, according to prior reporting. So the headline number ($650M) is actually smaller than the round before it, which is the first hint that this raise is about operations, not hype.</p> <h2> The real story: inference is the prize now </h2> <p>For three years the AI infrastructure narrative was about <em>training</em>, the giant clusters that build a model. Groq's raise is one more data point that the durable money is moving to <em>inference</em>, the cost of running that model billions of times a day once it exists. Infinitum's John Yetimoglu put the thesis bluntly: "We believe inference will become the largest infrastructure market in technology" (Groq newsroom).</p> <p>That matters because inference economics are the opposite of training economics. Training is a capital event; inference is a utility bill that never stops. Every chatbot reply, every agent step, every retrieval call is an inference cost, and as products move from demos to production, that line item is what decides whether an AI feature is profitable. We covered the brutal version of this in <a href="https://ai.crescevo.com/openai-losses-increased-nearly-8x-in-202/" rel="noopener noreferrer">OpenAI's losses increasing nearly 8x</a>: the bill for serving models is the story of the industry right now.</p> <h2> The substance: what a "neocloud" actually sells </h2> <p>Groq is positioning as an inference "neocloud", a specialized cloud that does one thing (fast, cheap token generation) rather than competing with AWS across everything. Its pitch rests on the LPU, a chip designed specifically for the sequential math of inference rather than the parallel math of training. The claimed advantages are latency and cost per token, not raw model quality.</p> <p>The numbers Groq offers as proof of scale: more than 5 million developers on the platform and "trillions of tokens" processed weekly (Groq newsroom). Those are usage figures, not revenue, and they're worth reading skeptically, a large share of developer signups on any free inference tier never become paying volume. But the 200-megawatt target by end of 2027 is the figure that actually signals intent: power, not chips, is the binding constraint on inference at scale, and committing to 200 MW is a commitment to a real, physical buildout.</p> <h2> Who wins, who loses </h2> <ul> <li> <strong>Wins , buyers of inference.</strong> A funded, independent inference cloud is good news for any team choosing a provider. More credible competition to Nvidia-on-the-hyperscalers means pricing pressure on tokens.</li> <li> <strong>Wins , Nvidia, oddly.</strong> Nvidia got Groq's IP and talent for $20B in December and announced an LPX platform at GTC that incorporates Groq's inference technology (per reporting). Now an independent Groq will deploy that lineage at scale. Nvidia benefits whether you buy from it or from Groq.</li> <li> <strong>Loses , pure-play training-chip narratives.</strong> Capital voting for inference is capital not voting for the next training-cluster startup.</li> <li> <strong>Uncertain , Groq's own margins.</strong> Running a neocloud is a low-margin, capital-heavy infrastructure business, closer to a utility than to software. Raising $650M to fund 200 MW is a sign of how expensive this path is.</li> </ul> <h2> The non-obvious angle most coverage missed </h2> <p>Here is the part the funding headlines skip. Groq is rebuilding as an inference cloud at least partly on the same technology it licensed to Nvidia, and it's deploying Nvidia's new LPX system, which itself incorporates Groq's inference IP, inside its own data centers (per reporting). In other words, the company sold its crown jewels to its largest competitor, then raised money to operate a cloud that runs on a productized version of those same jewels.</p> <p>That is clever and fragile at once. Clever, because Groq converts a one-time $20B licensing windfall into an operating business without having to keep winning the chip-design arms race against Nvidia. Fragile, because a neocloud whose hardware roadmap now partly depends on its biggest rival has surrendered the one thing, proprietary silicon, that made it special. The bet the new investors are making is not "Groq builds the best chip." It's "inference demand is so large that even a Groq without its founder, and partly dependent on Nvidia, is a good business." That's a bet on the <em>market</em>, not the company, and it tells you exactly how big the inference market has become.</p> <h2> What this means for you </h2> <ul> <li> <strong>If you ship AI features:</strong> treat inference as a sourcing decision, not a default. A capitalized independent like Groq is leverage in your pricing conversations with incumbents. Benchmark cost-per-token <em>and</em> latency on your real workloads before committing.</li> <li> <strong>If you're planning AI budget:</strong> model inference as a recurring utility that grows with usage, not a fixed build cost. The companies winning on AI margins are the ones who treated token spend as a first-class line item early.</li> <li> <strong>If you're an investor or operator watching the space:</strong> the tell to track is power, not chips. Watch who actually lands megawatts (Groq's 200 MW by 2027), because that, not benchmark slides, is what constrains real inference supply.</li> </ul> <h2> Frequently asked questions </h2> <h3> How much did Groq raise and who led the round? </h3> <p>Groq raised $650 million, announced June 22, 2026, led by Disruptive and the hedge fund Infinitum, with existing investors reinvesting, according to Groq's newsroom.</p> <h3> What was Groq's valuation? </h3> <p>The new round's valuation was not disclosed. Groq was last reported at a $6.9 billion valuation in a $750 million round in September 2025, up from $2.8 billion in August 2024, per prior reporting.</p> <h3> Didn't Nvidia already take over Groq? </h3> <p>Not as an acquisition. In December 2025 Nvidia struck a roughly $20 billion deal to license Groq's LPU technology and hire senior staff, including founder Jonathan Ross, per reporting. Nvidia stated it had not acquired Groq. Groq remains an independent company and has since brought in new leadership.</p> <h3> What is an inference "neocloud"? </h3> <p>A cloud specialized for running (not training) AI models, optimized for low-latency, low-cost token generation. Groq sells access to its LPU-based inference rather than competing as a general-purpose cloud.</p> <h2> Sources </h2> <ul> <li><a href="https://groq.com/newsroom/groq-raises-usd650m-to-scale-its-ai-inference-cloud-business?ref=ai.crescevo.com" rel="noopener noreferrer">Groq newsroom , Groq Raises $650M to Scale Its AI Inference Cloud Business</a></li> <li><a href="https://techcrunch.com/2026/05/29/after-nvidias-20b-not-acqui-hire-ai-chip-startup-groq-reportedly-raising-650m/?ref=ai.crescevo.com" rel="noopener noreferrer">TechCrunch , After Nvidia's $20B not-acqui-hire, Groq reportedly raising $650M</a></li> <li><a href="https://siliconangle.com/2026/06/22/inference-chip-startup-groq-raises-650m-grow-cloud-platform/?ref=ai.crescevo.com" rel="noopener noreferrer">SiliconANGLE , Inference chip startup Groq raises $650M to grow its cloud platform</a></li> <li><a href="https://www.pymnts.com/news/investment-tracker/2026/groq-seeks-650-million-amid-shift-to-ai-inference-neocloud-business/?ref=ai.crescevo.com" rel="noopener noreferrer">PYMNTS , Groq Seeks $650 Million Amid Shift to AI Inference Neocloud Business</a></li> </ul> <h2> Related reading </h2> <ul> <li><a href="https://ai.crescevo.com/real-time-llm-inference-on-standard-gpus/" rel="noopener noreferrer">Real-time LLM Inference on Standard GPUs: 3k tokens/s per request</a></li> <li><a href="https://ai.crescevo.com/duckduckgo-search-saw-28-more-visits-aft/" rel="noopener noreferrer">DuckDuckGo search saw 28% more visits after Google said people love AI mode</a></li> <li><a href="https://ai.crescevo.com/steve-wozniak-cheered-after-telling-stud/" rel="noopener noreferrer">Steve Wozniak cheered after telling students they have AI – actual intelligence</a></li> </ul> <p><em>Originally published on <a href="https://ai.crescevo.com/groq-650m-inference-cloud-raise/" rel="noopener noreferrer">AI at Crescevo</a> — subscribe free for more.</em></p> ai machinelearning llm programming ironbyte-rgb Fri, 26 Jun 2026 17:00:15 +0000 https://dev.to/crescevo/chatgpt-just-lost-its-majority-the-real-story-is-ads-not-decline-a0g https://dev.to/crescevo/chatgpt-just-lost-its-majority-the-real-story-is-ads-not-decline-a0g <h2> TL;DR </h2> <ul> <li>ChatGPT's share of global AI-assistant users fell to <strong>46.4%</strong> by the end of May 2026 , its first time below 50% , per Sensor Tower's State of AI 2026 report (released June 16).</li> <li>It is still #1 by raw scale: <strong>1.1 billion+ monthly users</strong> vs Gemini's 662M (27.7%) and Claude's 245M (10.3%), per Sensor Tower.</li> <li>This is not decline , it's the market outgrowing one app. H1 2026 saw <strong>2.3 billion</strong> AI-app downloads and <strong>$4.2B</strong> in spend (up from $1.83B a year earlier).</li> <li>The under-covered story: OpenAI is now serving <strong>ads to ~17% of daily ChatGPT users</strong> (up from a February test), with Walmart, Target, and Costco shopping integrations.</li> </ul> <p>ChatGPT just lost its majority for the first time since it created the category in 2022. But "ChatGPT falls below 50%" is the wrong headline to fixate on , the app added users faster than ever and still doubles its nearest rival. The two things that actually matter for anyone building on or marketing through AI are quieter: the assistant layer is fragmenting into a real multi-model market, and OpenAI has started turning that attention into an ad business.</p> <h2> What the numbers actually say </h2> <p>Per Sensor Tower, ChatGPT's "true audience" share (measured across desktop, mobile apps, and mobile web) went 65.3% in December 2024 → 52.8% in December 2025 → 46.4% by May 2026, crossing below 50% in March. Yet in absolute terms it crossed 1.1 billion monthly active users in May , the fastest app in history to that mark, ahead of TikTok, YouTube, and Instagram. Both things are true: ChatGPT is bigger than ever <em>and</em> a minority of a much larger market.</p> <p>The share moved because the field grew. Gemini sits second at 27.7% (662M users), Claude third at 10.3% (245M), and the rest , Meta AI, Grok, Perplexity, DeepSeek , collectively ~15.6%. The standout is Claude: Sensor Tower clocked its true audience up 452% year-over-year in May, with US share climbing from 4.4% to nearly 14%.</p> <h2> The real story #1: the assistant layer is now multi-model </h2> <p>For two years "AI strategy" meant "ChatGPT strategy." That's over. A market where the leader holds 46% and two credible challengers hold a combined ~38% is a market you can no longer address through one model. If you're optimizing content for AI visibility, building on an assistant API, or thinking about where your customers ask questions, the answer is now plural , Gemini and Claude are not rounding errors. Single-model dependence is becoming a real risk, both for cost and for reach.</p> <h2> The real story #2: ChatGPT is becoming an ad network </h2> <p>OpenAI started testing ads in ChatGPT in February and has scaled steadily , by May, an average of 17% of daily users were seeing them, per Sensor Tower. The largest advertiser categories so far are software and shopping, then media/entertainment and food/dining, and OpenAI has wired in shopping with Walmart, Target, and Costco. That is the beginning of a new, enormous ad surface , one that sits inside the answer, not beside it. For marketers, "being recommended by ChatGPT" is shifting from a pure-organic game toward a paid one. For users, the ad-free era that drove adoption is ending, which itself adds switching friction (a tailwind for Gemini and Claude).</p> <h2> The trust variable nobody prices in </h2> <p>One data point worth holding onto: when OpenAI announced a US Department of Defense deal in February 2026, Sensor Tower recorded a measurable spike in ChatGPT uninstalls and a matching surge in Claude downloads. In a multi-model market with low switching cost, brand values move share. That's new , and it favors whichever provider a given audience trusts most.</p> <h2> What this means for you </h2> <ul> <li> <strong>Stop optimizing for one model.</strong> Treat AI visibility as multi-model , test how Gemini and Claude answer your category, not just ChatGPT.</li> <li> <strong>Plan for paid AI placement.</strong> Ads inside ChatGPT answers are coming at scale; budget and positioning will follow, the way they did for search.</li> <li> <strong>If you build products:</strong> avoid single-provider lock-in. Pricing, availability, and trust now vary enough across ChatGPT/Gemini/Claude that portability is leverage.</li> <li> <strong>Watch the OpenAI S-1.</strong> OpenAI filed confidentially on June 8, 2026; the ad ramp is partly a pre-IPO revenue story, so expect it to accelerate.</li> </ul> <h2> Frequently asked questions </h2> <h3> Did ChatGPT actually lose users? </h3> <p>No. ChatGPT grew to 1.1 billion+ monthly users by May 2026 , a record. Its <em>share</em> fell to 46.4% because the overall AI-assistant market grew faster, led by Gemini and Claude.</p> <h3> Who is gaining the most? </h3> <p>Claude. Sensor Tower reported its true audience up 452% year-over-year in May, with US market share rising from 4.4% to nearly 14%. Gemini remains the #2 overall at 27.7%.</p> <h3> Is ChatGPT really showing ads now? </h3> <p>Yes. OpenAI began testing ads in February 2026 and, per Sensor Tower, ~17% of daily users were being served ads by May, with shopping integrations including Walmart, Target, and Costco.</p> <h3> What should marketers do about it? </h3> <p>Go multi-model (test Gemini and Claude, not just ChatGPT) and prepare for paid placement inside AI answers as ChatGPT's ad surface scales.</p> <h2> Sources </h2> <ul> <li><a href="https://techcrunch.com/2026/06/16/chatgpts-market-share-slips-below-50-for-first-time/?ref=ai.crescevo.com" rel="noopener noreferrer">TechCrunch , ChatGPT's market share slips below 50% for the first time</a></li> <li><a href="https://www.chaincatcher.com/en/article/2271917?ref=ai.crescevo.com" rel="noopener noreferrer">Sensor Tower State of AI 2026 report (summary)</a></li> <li><a href="https://www.fastcompany.com/91560276/chatgpt-loses-ground-gemini-claude-below-50-percent-market-share?ref=ai.crescevo.com" rel="noopener noreferrer">Fast Company , ChatGPT loses ground to Gemini and Claude</a></li> </ul> <h2> Related reading </h2> <ul> <li><a href="https://ai.crescevo.com/real-time-llm-inference-on-standard-gpus/" rel="noopener noreferrer">Real-time LLM Inference on Standard GPUs: 3k tokens/s per request</a></li> <li><a href="https://ai.crescevo.com/claude-fable-5-benchmark-vs-real-world-coding/" rel="noopener noreferrer">Claude Fable 5 Scores 95% on Its Own Benchmark and 19% on Real Security Work. The Gap Is the Lesson.</a></li> <li><a href="https://ai.crescevo.com/592-signals-6-niches-what-breaks-through-2026/" rel="noopener noreferrer">I Tracked 592 Signals Across 6 B2B Niches for 3 Weeks. AI Won — But Not How You'd Expect</a></li> </ul> <p><em>Originally published on <a href="https://ai.crescevo.com/chatgpt-below-50-percent-market-share/" rel="noopener noreferrer">AI at Crescevo</a> — subscribe free for more.</em></p> ai machinelearning llm programming ironbyte-rgb Fri, 26 Jun 2026 16:00:15 +0000 https://dev.to/crescevo/gemini-35-pro-and-the-announcement-to-shipping-gap-costing-google-49m1 https://dev.to/crescevo/gemini-35-pro-and-the-announcement-to-shipping-gap-costing-google-49m1 <h2> TL;DR </h2> <ul> <li>Gemini 3.5 Pro , announced at Google I/O on May 19, 2026 , entered its general-availability window on <strong>June 23</strong>, but as of the latest reporting it remains in <strong>limited preview for Vertex AI enterprise customers only</strong>, not the public Gemini app or AI Studio.</li> <li>Headline spec: a <strong>2-million-token context window</strong> , the largest in any production frontier model, reportedly ~10x GPT-5 and ~16x Claude's current limit.</li> <li> <strong>Deep Think</strong>, its extended-reasoning mode, is gated to Google's <strong>$250/month Ultra</strong> tier , the most expensive consumer AI subscription on the market.</li> <li>Pricing (~$15/$60 per million tokens) and benchmark gains remain <strong>Google's claims, not verified</strong>; prediction markets put the odds of a public release by June 30 at only ~50-55%.</li> </ul> <p>Google's next flagship is technically "launching" this week, and that's exactly the problem. Gemini 3.5 Pro has genuinely impressive announced specs , a 2M-token context window no competitor matches , but it has slipped repeatedly since I/O, and the real story isn't the model. It's the widening gap between what Google <em>announces</em> and what Google <em>ships</em>.</p> <h2> What's actually shipping (and what isn't) </h2> <p>Sundar Pichai told the I/O audience on May 19 to "give us until next month." It's now late June, and Gemini 3.5 Pro is still in limited preview for select Vertex AI enterprise customers; it has not reached the public Gemini app, AI Studio, or the consumer subscription. Until it does, <strong>Gemini 3.1 Pro remains Google's GA flagship.</strong> Treat the specs below as Google's positioning , the verified numbers only arrive with the model card at true GA.</p> <h2> The 2-million-token context window , the one real differentiator </h2> <p>The standout, if it ships as described, is the context window: 2 million tokens, double Gemini 3.5 Flash and the largest deployed in any production frontier model. Reporting frames it as ~10x GPT-5's window and ~16x Claude's current production limit. This is not a benchmark-bragging-rights number , it's a capability unlock. A 2M window means reasoning over an entire codebase, a full data room, or months of conversation history in a single request, without the retrieval-and-stitching gymnastics that long-context workflows require today. For enterprise document and code use cases, that is the feature that actually matters.</p> <h2> Deep Think, and the $250/month question </h2> <p>Deep Think is Google's "think before answering" mode , a hidden-scratchpad reasoning approach similar to OpenAI's o-series. The notable part is the packaging: Deep Think is gated exclusively to the $250/month Ultra tier, making it the priciest consumer-facing AI subscription by a wide margin. Google is betting a slice of users will pay flagship-software money for frontier reasoning. Whether that tier finds an audience is its own open question.</p> <h2> The real story: announce-vs-ship is Google's recurring tax </h2> <p>Google keeps pre-announcing frontier capability it then struggles to ship on time , and competitors keep filling the gap. While 3.5 Pro sits in preview, OpenAI and Anthropic ship into general availability. The cost isn't just a slipped date; it's that "Gemini 3.5 Pro" has been a talking point for over a month with no public model behind it, which trains developers to build on what's actually available (3.1 Pro, or rivals) rather than wait. In a market where switching is cheap, the announcement-to-availability gap is where Google leaks momentum , and this launch is being read as a test of whether it can close it.</p> <h2> What this means for you </h2> <ul> <li> <strong>Don't architect on unreleased specs.</strong> The 2M window and pricing are Google's claims until the GA model card. Build on what you can call today; pilot 3.5 Pro only once it's actually in AI Studio/Vertex GA.</li> <li> <strong>If long context is your bottleneck:</strong> 3.5 Pro is the one to evaluate the day it ships , whole-codebase and whole-corpus reasoning is a genuine step change worth testing on your real workloads.</li> <li> <strong>Price the reasoning tier carefully.</strong> Deep Think at $250/mo is premium; validate that the quality delta justifies it for your use case before committing seats.</li> <li> <strong>Hedge providers.</strong> Until Gemini's GA cadence is reliable, single-vendor dependence carries schedule risk.</li> </ul> <h2> Frequently asked questions </h2> <h3> Is Gemini 3.5 Pro available now? </h3> <p>Not broadly. As of late June 2026 it's in limited preview for select Vertex AI enterprise customers; it has not reached the public Gemini app, AI Studio, or the consumer subscription. Gemini 3.1 Pro remains the GA flagship.</p> <h3> What is the 2-million-token context window good for? </h3> <p>Processing very large inputs , entire codebases, large document sets, long histories , in one request. Reporting puts it at roughly 10x GPT-5 and 16x Claude's current production limit, the largest in any production frontier model.</p> <h3> What does Deep Think cost? </h3> <p>It's gated to Google's Ultra tier at $250/month, the most expensive consumer AI subscription currently on the market.</p> <h3> Are the benchmarks and pricing confirmed? </h3> <p>No. Pricing (~$15/$60 per million tokens) and benchmark gains are reported expectations, not verified facts. Official numbers arrive with the GA model card.</p> <h2> Sources </h2> <ul> <li><a href="https://www.techtimes.com/articles/317919/20260606/google-gemini-35-pro-nears-june-launch-2-million-token-context-deep-think-reasoning.htm?ref=ai.crescevo.com" rel="noopener noreferrer">Tech Times , Gemini 3.5 Pro nears June launch with 2M context and Deep Think</a></li> <li><a href="https://fourweekmba.com/gemini-3-5-pro-2-million-context-deep-think-launch/?ref=ai.crescevo.com" rel="noopener noreferrer">FourWeekMBA , Gemini 3.5 Pro is days away</a></li> <li><a href="https://enterprisedna.co/resources/news/google-gemini-35-pro-nears-launch-deep-think-2m-tokens-2026/?ref=ai.crescevo.com" rel="noopener noreferrer">Enterprise DNA , Gemini 3.5 Pro nears launch</a></li> </ul> <h2> Related reading </h2> <ul> <li><a href="https://ai.crescevo.com/claude-fable-5-benchmark-vs-real-world-coding/" rel="noopener noreferrer">Claude Fable 5 Scores 95% on Its Own Benchmark and 19% on Real Security Work. The Gap Is the Lesson.</a></li> <li><a href="https://ai.crescevo.com/duckduckgo-search-saw-28-more-visits-aft/" rel="noopener noreferrer">DuckDuckGo search saw 28% more visits after Google said people love AI mode</a></li> <li><a href="https://ai.crescevo.com/the-ai-product-managers-checklist-for-shipping-responsibly/" rel="noopener noreferrer">The AI Product Manager's Checklist for Shipping Responsibly</a></li> </ul> <p><em>Originally published on <a href="https://ai.crescevo.com/gemini-3-5-pro-general-availability/" rel="noopener noreferrer">AI at Crescevo</a> — subscribe free for more.</em></p> ai machinelearning llm programming ironbyte-rgb Wed, 24 Jun 2026 19:00:15 +0000 https://dev.to/crescevo/real-time-llm-inference-on-standard-gpus-3k-tokenss-per-request-38d6 https://dev.to/crescevo/real-time-llm-inference-on-standard-gpus-3k-tokenss-per-request-38d6 <h2> TL;DR </h2> <ul> <li>Real-time LLM inference on standard GPUs can reach 3k tokens/s per request</li> <li>Optimizing the whole software stack with architecture/engine/kernel co-design is crucial for fast inference</li> <li>Standard datacenter GPU hardware has a higher decoding-speed ceiling than current inference stacks expose</li> <li>The limiting factor is existing inference software stacks not optimized for single-request decoding speed</li> </ul> <p>The primary source article explains why optimizing for single-request LLM decoding speed is important for AI agents and how standard datacenter GPUs can achieve this speed. According to the article, the key to reaching this speed is co-designing the model architecture, runtime, and low-level GPU code as a single latency-optimized pipeline. This approach allows for extremely fast single-request decoding without the need for proprietary silicon.</p> <h2> What the data shows </h2> <p>The article highlights that existing inference software stacks are not optimized for single-request decoding speed, which is a memory-bandwidth maximization problem rather than a FLOPS one. The data shows that standard datacenter GPUs can achieve a much higher decoding-speed ceiling than current inference stacks expose, but this requires optimizing the whole software stack. The article also notes that inference benchmarks typically conflate three quantities: aggregate throughput, time to first token, and decode speed per request. Decode speed per request is the metric that matters for AI agents, as it governs every long serial interaction.</p> <h2> What this means for AI readers </h2> <p>For AI readers, the ability to achieve real-time LLM inference on standard GPUs means that they can build more responsive and interactive products. As the article notes, if an agent needs to generate 50,000 tokens in a workflow, 100 tokens/s is roughly eight minutes, while 3,000 tokens/s is under twenty seconds. This difference can significantly change the product that can be built. The article also explains that agentic software engineering is a sequential loop, and the generation-heavy steps set the loop rate. Therefore, faster decode speeds can lead to more efficient and effective AI agents.</p> <h2> What to do right now </h2> <p>The article invites readers to test the speed of their 2B coding model in their live coding playground: playground.kog.ai. This allows readers to experience the fast single-request decoding speed firsthand. The article also notes that the 2B coding model is small and not a frontier model, but it is still quite capable when fine-tuned for specific software engineering tasks.</p> <h2> Bottom line </h2> <p>The primary source article demonstrates that real-time LLM inference on standard GPUs is possible, achieving 3k tokens/s per request. This is made possible by optimizing the whole software stack with architecture/engine/kernel co-design. The article highlights the importance of decode speed per request for AI agents and how standard datacenter GPUs can achieve this speed without the need for proprietary silicon. By co-designing the model architecture, runtime, and low-level GPU code, developers can build more responsive and interactive products.</p> <h2> Frequently asked questions </h2> <h3> Q: What is the current limitation of existing inference software stacks? </h3> <p>The current limitation is that existing inference software stacks are not optimized for single-request decoding speed, which is a memory-bandwidth maximization problem rather than a FLOPS one.</p> <h3> Q: What is the key to achieving fast single-request decoding speed? </h3> <p>The key is co-designing the model architecture, runtime, and low-level GPU code as a single latency-optimized pipeline.</p> <h3> Q: How can readers experience the fast single-request decoding speed firsthand? </h3> <p>Readers can test the speed of the 2B coding model in the live coding playground: playground.kog.ai.</p> <h3> Q: What is the significance of achieving 3k tokens/s per request for AI agents? </h3> <p>Achieving 3k tokens/s per request can significantly change the product that can be built, as it enables more responsive and interactive products, and faster decode speeds can lead to more efficient and effective AI agents.</p> <h2> Sources </h2> <ul> <li><a href="https://blog.kog.ai/real-time-llm-inference-on-standard-gpus-3-000-tokens-s-per-request/?ref=ai.crescevo.com" rel="noopener noreferrer">https://blog.kog.ai/real-time-llm-inference-on-standard-gpus-3-000-tokens-s-per-request/</a></li> </ul> <h2> Related reading </h2> <ul> <li><a href="https://ai.crescevo.com/claude-fable-5-benchmark-vs-real-world-coding/" rel="noopener noreferrer">Claude Fable 5 Scores 95% on Its Own Benchmark and 19% on Real Security Work. The Gap Is the Lesson.</a></li> <li><a href="https://ai.crescevo.com/if-you-re-an-llm-2/" rel="noopener noreferrer">If you're an LLM</a></li> <li><a href="https://ai.crescevo.com/the-evaluation-framework-that-actually-predicts-llm-production-performance/" rel="noopener noreferrer">The Evaluation Framework That Actually Predicts LLM Production Performance</a></li> </ul> <p><em>Originally published on <a href="https://ai.crescevo.com/real-time-llm-inference-on-standard-gpus/" rel="noopener noreferrer">AI at Crescevo</a> — subscribe free for more.</em></p> ai machinelearning llm programming ironbyte-rgb Tue, 23 Jun 2026 19:00:15 +0000 https://dev.to/crescevo/ponytail-make-your-ai-agent-think-like-the-laziest-senior-dev-in-the-room-34h0 https://dev.to/crescevo/ponytail-make-your-ai-agent-think-like-the-laziest-senior-dev-in-the-room-34h0 <h2> TL;DR </h2> <ul> <li>Ponytail reduces code by ~54% on average, with a maximum reduction of ~94% in certain cases.</li> <li>It also reduces costs by ~20% and time by ~27%, while maintaining 100% safety.</li> <li>Ponytail achieves these results by making an AI agent think like a lazy senior developer, who writes minimal code to achieve the desired outcome.</li> <li>The results are based on a study of 12 feature tasks, using a real open-source repository and a headless Claude Code session.</li> </ul> <p>Ponytail is a technique that makes an AI agent think like a lazy senior developer, who writes minimal code to achieve the desired outcome. This approach has been shown to reduce code by ~54% on average, with a maximum reduction of ~94% in certain cases. The technique was tested on a real open-source repository, using a headless Claude Code session, and the results show significant reductions in code, cost, and time.</p> <h2> What the data shows </h2> <p>The data shows that Ponytail reduces code by ~54% on average, with a maximum reduction of ~94% in certain cases. For example, when asked to create a date picker, the AI agent without Ponytail installed flatpickr, wrote a wrapper component, added a stylesheet, and started a discussion about timezones. In contrast, the AI agent with Ponytail simply wrote ``. The study also shows that Ponytail reduces costs by ~20% and time by ~27%, while maintaining 100% safety.</p> <h2> What this means for AI readers </h2> <p>The results of the study have significant implications for AI readers, as they show that it is possible to make AI agents think like lazy senior developers, who write minimal code to achieve the desired outcome. This approach can help reduce the amount of code that needs to be written, which can in turn reduce costs and time. The study also shows that Ponytail is the only approach that cuts every metric, including code, cost, and time, while maintaining 100% safety.</p> <h2> What to do right now </h2> <p>To take advantage of Ponytail, readers can start by trying out the technique on their own AI agents. The study provides a reproducible example, using a headless Claude Code session and a real open-source repository. Readers can also review the full method, per-task tables, and limitations of the study, which are available in the benchmarks/results/2026-06-18-agentic.md file.</p> <h2> Bottom line </h2> <p>In summary, Ponytail is a technique that makes an AI agent think like a lazy senior developer, who writes minimal code to achieve the desired outcome. The results of the study show that Ponytail reduces code by ~54% on average, with a maximum reduction of ~94% in certain cases, while also reducing costs by ~20% and time by ~27%, and maintaining 100% safety.</p> <h2> Frequently asked questions </h2> <h3> Q: What is Ponytail? </h3> <p>Ponytail is a technique that makes an AI agent think like a lazy senior developer, who writes minimal code to achieve the desired outcome.</p> <h3> Q: How much code reduction can Ponytail achieve? </h3> <p>Ponytail can reduce code by ~54% on average, with a maximum reduction of ~94% in certain cases.</p> <h3> Q: Is Ponytail safe to use? </h3> <p>Yes, Ponytail maintains 100% safety, while also reducing code, cost, and time.</p> <h3> Q: Where can I find more information about Ponytail? </h3> <p>More information about Ponytail can be found on the primary source URL: <a href="https://github.com/DietrichGebert/ponytail" rel="noopener noreferrer">https://github.com/DietrichGebert/ponytail</a></p> <h2> Sources </h2> <ul> <li><a href="https://github.com/DietrichGebert/ponytail?ref=ai.crescevo.com" rel="noopener noreferrer">https://github.com/DietrichGebert/ponytail</a></li> </ul> <h2> Related reading </h2> <ul> <li><a href="https://ai.crescevo.com/i-think-anthropic-and-openai-have-found/" rel="noopener noreferrer">I think Anthropic and OpenAI have found product-market fit</a></li> <li><a href="https://ai.crescevo.com/about-llms-at-zig-days/" rel="noopener noreferrer">About LLMs at Zig Days</a></li> <li><a href="https://ai.crescevo.com/openai-losses-increased-nearly-8x-in-202/" rel="noopener noreferrer">OpenAI Losses Increased Nearly 8X in 2025, with Spending Hitting $34B</a></li> </ul> <p><em>Originally published on <a href="https://ai.crescevo.com/ponytail-make-your-ai-agent-think-like-t/" rel="noopener noreferrer">AI at Crescevo</a> — subscribe free for more.</em></p> ai machinelearning llm programming ironbyte-rgb Mon, 22 Jun 2026 19:00:15 +0000 https://dev.to/crescevo/trees-to-flows-and-back-unifying-decision-trees-and-diffusion-models-327l https://dev.to/crescevo/trees-to-flows-and-back-unifying-decision-trees-and-diffusion-models-327l <h2> TL;DR </h2> <ul> <li>The paper "Trees to Flows and Back: Unifying Decision Trees and Diffusion Models" establishes a mathematical correspondence between decision trees and diffusion models.</li> <li>The authors, Sai Niranjan Ramachandran and Suvrit Sra, introduce Global Trajectory Score Matching (GTSM) as a shared optimization principle.</li> <li>The work leads to two practical instantiations: \treeflow and \dsmtree, which achieve competitive generation quality and transfer hierarchical decision logic into neural networks.</li> <li>The paper was accepted in the Forty-Third International Conference on Machine Learning (ICML) 2026.</li> </ul> <p>The paper "Trees to Flows and Back: Unifying Decision Trees and Diffusion Models" by Sai Niranjan Ramachandran and Suvrit Sra presents a significant contribution to the field of machine learning. The authors establish a crisp mathematical correspondence between hierarchical decision trees and diffusion processes in appropriate limiting regimes. This unification reveals a shared optimization principle, Global Trajectory Score Matching (GTSM), for which gradient boosting is asymptotically optimal.</p> <h2> What the data shows </h2> <p>The paper provides evidence of the effectiveness of the proposed approach through two key practical instantiations: \treeflow and \dsmtree. \treeflow achieves competitive generation quality on tabular data with higher fidelity and a 2\times computational speedup. \dsmtree, a novel distillation method, transfers hierarchical decision logic into neural networks, matching teacher performance within 2\% on many benchmarks. The authors' work is supported by a 12-page main paper and a 68-page appendix.</p> <h2> What this means for ai readers </h2> <p>The unification of decision trees and diffusion models has significant implications for the field of artificial intelligence. The introduction of Global Trajectory Score Matching (GTSM) as a shared optimization principle provides a new perspective on the optimization of machine learning models. The practical instantiations of \treeflow and \dsmtree demonstrate the potential of this approach to improve the performance of machine learning models.</p> <h2> What to do right now </h2> <p>Readers interested in learning more about the paper can access the PDF of the paper titled "Trees to Flows and Back: Unifying Decision Trees and Diffusion Models" by Sai Niranjan Ramachandran and Suvrit Sra. The paper is available on the arXiv website and has been accepted in the Forty-Third International Conference on Machine Learning (ICML) 2026. Readers can also explore the authors' other works and research in the field of machine learning.</p> <h2> Bottom line </h2> <p>The paper "Trees to Flows and Back: Unifying Decision Trees and Diffusion Models" presents a significant contribution to the field of machine learning. The authors' work establishes a mathematical correspondence between decision trees and diffusion models, introducing Global Trajectory Score Matching (GTSM) as a shared optimization principle. The practical instantiations of \treeflow and \dsmtree demonstrate the potential of this approach to improve the performance of machine learning models.</p> <h2> Frequently asked questions </h2> <h3> Q: What is the title of the paper? </h3> <p>The title of the paper is "Trees to Flows and Back: Unifying Decision Trees and Diffusion Models".</p> <h3> Q: Who are the authors of the paper? </h3> <p>The authors of the paper are Sai Niranjan Ramachandran and Suvrit Sra.</p> <h3> Q: What is the main contribution of the paper? </h3> <p>The main contribution of the paper is the establishment of a mathematical correspondence between decision trees and diffusion models, introducing Global Trajectory Score Matching (GTSM) as a shared optimization principle.</p> <h3> Q: Where can I access the paper? </h3> <p>The paper is available on the arXiv website, and the PDF can be accessed through the link provided in the primary source URL: <a href="https://arxiv.org/abs/2605.00414" rel="noopener noreferrer">https://arxiv.org/abs/2605.00414</a>.</p> <h2> Sources </h2> <ul> <li><a href="https://arxiv.org/abs/2605.00414?ref=ai.crescevo.com" rel="noopener noreferrer">https://arxiv.org/abs/2605.00414</a></li> </ul> <h2> Related reading </h2> <ul> <li><a href="https://ai.crescevo.com/about-llms-at-zig-days/" rel="noopener noreferrer">About LLMs at Zig Days</a></li> <li><a href="https://ai.crescevo.com/openai-losses-increased-nearly-8x-in-202/" rel="noopener noreferrer">OpenAI Losses Increased Nearly 8X in 2025, with Spending Hitting $34B</a></li> <li><a href="https://ai.crescevo.com/spacex-buys-cursor-anysphere-60-billion-ai-coding/" rel="noopener noreferrer">SpaceX Is Paying $60 Billion for Cursor. The Tell Is That Cursor Is Losing.</a></li> </ul> <p><em>Originally published on <a href="https://ai.crescevo.com/trees-to-flows-and-back-unifying-decisio/" rel="noopener noreferrer">AI at Crescevo</a> — subscribe free for more.</em></p> ai machinelearning llm programming ironbyte-rgb Mon, 22 Jun 2026 15:00:15 +0000 https://dev.to/crescevo/top-remote-tech-jobs-this-week-june-22-2026-2458 https://dev.to/crescevo/top-remote-tech-jobs-this-week-june-22-2026-2458 <p>The best remote Tech jobs this week. All roles are fully remote. Updated June 22, 2026.</p> <h3> <strong>Steward — Members Don't Join for Dollar Value</strong> </h3> <p><em>teaching · finance · c · exec · golang</em></p> <p>There has been an idea rolling around the association space that members sign up for economic reasons.The gist of it is that you need to look at all your benefits and calculate a market price for each...</p> <p><a href="https://remoteok.com/remote-jobs/remote-members-dont-join-for-dollar-value-steward-1133822?ref=tech.crescevo.com" rel="noopener noreferrer"><strong>Apply →</strong></a></p> <h3> <strong>Steward — Differentiating Membership Offers</strong> </h3> <p><em>teaching · finance · c · exec · golang</em></p> <p>"They're not the only game in town anymore," a peer recently told me, "I bet they're having to learn how to compete like most private companies do."We were discussing associations and the challenges o...</p> <p><a href="https://remoteok.com/remote-jobs/remote-differentiating-membership-offers-steward-1133823?ref=tech.crescevo.com" rel="noopener noreferrer"><strong>Apply →</strong></a></p> <h3> <strong>Mississippi Army National Guard Biloxi Office — CYBER</strong> </h3> <p><em>teaching · finance · c · exec · golang</em></p> <p>In the cyber realm, anonymous attacks occur at near-light speed every day. Here, the Guard's growing cyber force fights on the frontlines of this digital domain. Cyber Soldiers are trained to execute...</p> <p><a href="https://remoteok.com/remote-jobs/remote-cyber-mississippi-army-national-guard-biloxi-office-1133824?ref=tech.crescevo.com" rel="noopener noreferrer"><strong>Apply →</strong></a></p> <h3> <strong>OWL Services — Data Analyst Data Warehouse Analytics and BI</strong> </h3> <p><em>analyst · design · sys admin · infosec · education</em></p> <p>Who We AreOWL Services is the premier sales, installation, program management and service provider to retail, commercial, fleet, aviation and marine, and emergency power generation companies across th...</p> <p><a href="https://remoteok.com/remote-jobs/remote-data-analyst-data-warehouse-analytics-and-bi-owl-services-1133828?ref=tech.crescevo.com" rel="noopener noreferrer"><strong>Apply →</strong></a></p> <h3> <strong>Ross Stores, Inc. — Assistant Store Manager</strong> </h3> <p><em>sys admin · technical · supervisor · customer support · testing</em></p> <p>Our values start with our people, join a team that values you!We are the nation’s largest off-price retailer with over 2,200 stores, and a strong track record of success and growth. Our focus has al...</p> <p><a href="https://remoteok.com/remote-jobs/remote-assistant-store-manager-ross-stores-inc-1133775?ref=tech.crescevo.com" rel="noopener noreferrer"><strong>Apply →</strong></a></p> <h3> <strong>Eaton — Manufacturing Operator I</strong> </h3> <p><em>sys admin · technical · supervisor · customer support · testing</em></p> <p>Lo Que HarásFUNCION PRINCIPAL Responsable de Operación y “set-up” de máquinas menos complejas y/o de apoyo a mas complejas y de ensamblaje, ajuste, y montaje de subconjuntos y conjuntos estánd...</p> <p><a href="https://remoteok.com/remote-jobs/remote-manufacturing-operator-i-eaton-1133763?ref=tech.crescevo.com" rel="noopener noreferrer"><strong>Apply →</strong></a></p> <h3> <strong>Haystack — Scrum Master</strong> </h3> <p><em>design · video · exec · content writing · digital nomad</em></p> <p>We're hiring on behalf of a Haystack partner!The Role Coach a development team on Agile best practices, focusing on sprint goals to ensure project success. Apply Scrum principles to resolve routine pr...</p> <p><a href="https://remoteok.com/remote-jobs/remote-scrum-master-haystack-1133814?ref=tech.crescevo.com" rel="noopener noreferrer"><strong>Apply →</strong></a></p> <h3> <strong>Talentoma — Project Assistant</strong> </h3> <p><em>analyst · design · sys admin · infosec · education</em></p> <p>Job Title: Remote Project AssistantMonthly Pay: $5,400 – $6,900Summary:The Remote Project Assistant provides administrative and operational support to project teams by coordinating schedules, mainta...</p> <p><a href="https://remoteok.com/remote-jobs/remote-project-assistant-talentoma-1133838?ref=tech.crescevo.com" rel="noopener noreferrer"><strong>Apply →</strong></a></p> <h3> <strong>AltaML — Growth Product Manager</strong> </h3> <p><em>product manager · marketing · exec · design · saas</em></p> <p>About the Role<br> Most product managers learn about customers secondhand, through tickets, surveys, and sales calls. This role is different.<br> At Brilliant Harvest, our customers are heavy equipment dealer...</p> <p><a href="https://remoteok.com/remote-jobs/remote-growth-product-manager-altaml-1133729?ref=tech.crescevo.com" rel="noopener noreferrer"><strong>Apply →</strong></a></p> <h3> <strong>RG&amp;T Solutions — No Experience</strong> </h3> <p><em>analyst · design · sys admin · infosec · education</em></p> <p>We are seeking motivated individuals for a Remote Part-Time Position based in El Dorado, Kansas, offering competitive pay rates of £18/hr - £32/hr. This opportunity is ideal for those looking to ear...</p> <p><a href="https://remoteok.com/remote-jobs/remote-no-experience-rgampt-solutions-1133829?ref=tech.crescevo.com" rel="noopener noreferrer"><strong>Apply →</strong></a></p> <h3> <strong>RG&amp;T Solutions — Data Entry Clerk</strong> </h3> <p><em>analyst · design · sys admin · infosec · education</em></p> <p>We are seeking motivated Remote Work From Home Data Entry Clerk - Part Time Panelists located in Madison, Mississippi. This flexible position offers the opportunity to contribute to various data colle...</p> <p><a href="https://remoteok.com/remote-jobs/remote-data-entry-clerk-rgampt-solutions-1133830?ref=tech.crescevo.com" rel="noopener noreferrer"><strong>Apply →</strong></a></p> <h3> <strong>RG&amp;T Solutions — Administrative Assistant Research Panel</strong> </h3> <p><em>analyst · design · sys admin · infosec · education</em></p> <p>We are seeking a motivated Part-Time Research Panel - Remote Work At Home (Administrative Assistant Welcome) to join our team in San Carlos, California. This role offers flexible hours and the opportu...</p> <p><a href="https://remoteok.com/remote-jobs/remote-administrative-assistant-research-panel-rgampt-solutions-1133835?ref=tech.crescevo.com" rel="noopener noreferrer"><strong>Apply →</strong></a></p> <p>Subscribe to the Tech Brief for weekly job picks and industry signals.</p> <p><em>Originally published on <a href="https://tech.crescevo.com/remote-tech-jobs-2026-06-22/" rel="noopener noreferrer">Tech at Crescevo</a> — subscribe free for more.</em></p> programming devops webdev tech ironbyte-rgb Mon, 22 Jun 2026 14:00:15 +0000 https://dev.to/crescevo/top-remote-ai-ml-jobs-this-week-june-22-2026-152n https://dev.to/crescevo/top-remote-ai-ml-jobs-this-week-june-22-2026-152n <p>The best remote AI &amp; ML jobs this week. All roles are fully remote. Updated June 22, 2026.</p> <h3> <strong>RECSEEKERS — Operations Administrator</strong> </h3> <p><em>hr · virtual assistant · medical · customer support · amazon</em></p> <p>EDTECH OPERATIONS ADMINISTRATOR<br> FULLY REMOTE | $75,000 - $110,000 BASE SALARY + BONUS +…See this and similar jobs on LinkedIn.Please mention the word <strong>RETRACTABLE</strong> and tag RNS43OC4xODguMjE= when a...</p> <p><a href="https://remoteok.com/remote-jobs/remote-operations-administrator-recseekers-1133675?ref=ai.crescevo.com" rel="noopener noreferrer"><strong>Apply →</strong></a></p> <h3> <strong>Swoon — Graphic Designer</strong> </h3> <p><em>hr · virtual assistant · medical · customer support · amazon</em></p> <p>Location: Fully RemoteDuration: 6-Month Contract with Potential ExtensionSchedule: Monday–Friday | Standard Business HoursAbout the OpportunityThis role is ideal for a highly creative presentation d...</p> <p><a href="https://remoteok.com/remote-jobs/remote-graphic-designer-swoon-1133676?ref=ai.crescevo.com" rel="noopener noreferrer"><strong>Apply →</strong></a></p> <h3> <strong>Blackstone Talent Group — Scrum Master</strong> </h3> <p><em>hr · virtual assistant · medical · customer support · amazon</em></p> <p>Scrum Master (Contractor) — Job DescriptionPosition Details:Location: Dallas, Tx (Remote)Type: 6 Month Contract - NO C2CResponsibilities: Support Agile delivery for assigned team(s) under guidance o...</p> <p><a href="https://remoteok.com/remote-jobs/remote-scrum-master-blackstone-talent-group-1133679?ref=ai.crescevo.com" rel="noopener noreferrer"><strong>Apply →</strong></a></p> <h3> <strong>Crypto.com — Social Media Manager</strong> </h3> <p><em>hr · virtual assistant · medical · customer support · amazon</em></p> <p>The Social Media Manager will be the driving force behind Crypto.com’s digital presence and community engagement. You will conceptualize, implement, and manage high-impact social strategies for...</p> <p><a href="https://remoteok.com/remote-jobs/remote-social-media-manager-crypto-com-1133674?ref=ai.crescevo.com" rel="noopener noreferrer"><strong>Apply →</strong></a></p> <h3> <strong>Zenius Corporation — Help Desk</strong> </h3> <p><em>hr · virtual assistant · medical · customer support · amazon</em></p> <p>Help Desk SpecialistLocation: RemotePosition SummaryWe are seeking a Help Desk Specialist to provide Tier 1/Tier 2 support for a Grants Management System (GMS). The selected candidate will serve as th...</p> <p><a href="https://remoteok.com/remote-jobs/remote-help-desk-zenius-corporation-1133680?ref=ai.crescevo.com" rel="noopener noreferrer"><strong>Apply →</strong></a></p> <h3> <strong>Archinect — Junior Project Designer</strong> </h3> <p><em>hr · virtual assistant · medical · customer support · amazon</em></p> <p>MHS Architecture EmployerHoboken, NJ, US (remote possible)Location:Full-timeTypeWed, Jun 17 '26Posted on:MHS Architecture is an award-winning interdisciplinary design office with a rich 40-year histor...</p> <p><a href="https://remoteok.com/remote-jobs/remote-junior-project-designer-archinect-1133677?ref=ai.crescevo.com" rel="noopener noreferrer"><strong>Apply →</strong></a></p> <h3> <strong>Shumba — Junior Market Specialist Analyst</strong> </h3> <p><em>crypto · blockchain</em></p> <p>Shumba Capital Ltd is an international company focused on digital assets, market research, and data-driven trading. We work with crypto markets and help our team grow from the ground up. We are fully...</p> <p><a href="https://remoteok.com/remote-jobs/remote-junior-market-specialist-analyst-shumba-1133023?ref=ai.crescevo.com" rel="noopener noreferrer"><strong>Apply →</strong></a></p> <h3> <strong>Gameplay Galaxy — Senior Backend Engineer</strong> </h3> <p><em>backend · senior · ai</em></p> <p>About the roleYou'll own the backend for a live mobile F2P game and the AI agentic platform that powers how we operate it. Feature releases, live ops, monetization, analytics pipeline, reliability. Se...</p> <p><a href="https://remoteok.com/remote-jobs/remote-senior-backend-engineer-gameplay-galaxy-1132512?ref=ai.crescevo.com" rel="noopener noreferrer"><strong>Apply →</strong></a></p> <h3> <strong>Storyteller — Client Delivery Manager</strong> </h3> <p><em>work from home · ai · saas · exec</em></p> <p>💸 Up to USD 80,000 per year, on a full time, contractor contract  🌎 Fully remote working! ✨ Opportunity to work close to major basketball and sports clients, including the live rhythms of...</p> <p><a href="https://remoteok.com/remote-jobs/remote-client-delivery-manager-storyteller-1132435?ref=ai.crescevo.com" rel="noopener noreferrer"><strong>Apply →</strong></a></p> <h3> <strong>Storyteller — Client Delivery Manager</strong> </h3> <p><em>work from home · ai · saas · exec</em></p> <p>🇨🇦 CAD 95,000 to CAD 110,000 salary, depending on experience, on a full time, permanent employment contract  🌎 Fully remote working anywhere in Canada! 🏖️ 33 Days Paid Leave and Bene...</p> <p><a href="https://remoteok.com/remote-jobs/remote-client-delivery-manager-storyteller-1132338?ref=ai.crescevo.com" rel="noopener noreferrer"><strong>Apply →</strong></a></p> <h3> <strong>CuraSenseAI — Data Annotator</strong> </h3> <p><em>data annotation · ai · artificial intelligence</em></p> <p>Role OverviewJoin a high-selectivity evaluation program focused on reviewing complex written…Please mention the word <strong>FAVOUR</strong> and tag RNS43OC4xODguMjE= when applying to show you read the job post...</p> <p><a href="https://remoteok.com/remote-jobs/remote-data-annotator-curasenseai-1132152?ref=ai.crescevo.com" rel="noopener noreferrer"><strong>Apply →</strong></a></p> <h3> <strong>Apprentus — Profesores de ELE para clases particulares</strong> </h3> <p><em>education · teaching · instructor · training · work from home</em></p> <p>PROFESORES/AS DE ESPAÑOL COMO LENGUA EXTRANJERA (ELE)Si buscas un proyecto educativo dinámico, flexible y en crecimiento, esta oportunidad es para ti.Buscamos docentes de Español como Lengua Extran...</p> <p><a href="https://remoteok.com/remote-jobs/remote-profesores-de-ele-para-clases-particulares-apprentus-1131998?ref=ai.crescevo.com" rel="noopener noreferrer"><strong>Apply →</strong></a></p> <p>Subscribe to the AI &amp; ML Brief for weekly job picks and industry signals.</p> <p><em>Originally published on <a href="https://ai.crescevo.com/remote-ai-jobs-2026-06-22/" rel="noopener noreferrer">AI at Crescevo</a> — subscribe free for more.</em></p> ai machinelearning llm programming ironbyte-rgb Mon, 22 Jun 2026 05:00:15 +0000 https://dev.to/crescevo/the-arm-takeover-is-complete-what-it-means-for-developers-1350 https://dev.to/crescevo/the-arm-takeover-is-complete-what-it-means-for-developers-1350 <h2> <strong>TL;DR</strong> </h2> <ul> <li>ARM-based silicon now dominates the high-performance laptop market with Apple's M1 series leading the way with roughly half the power draw of Intel's fastest laptop chips.</li> <li>AWS Graviton3 instances launched in 2022 offered 25% better compute performance and up to 60% better energy efficiency than comparable x86 instances at a 10–20% lower price point.</li> <li>Organizations running multi-million dollar AWS bills have moved 50–80% of their fleet to Graviton and are reporting meaningful cost reductions, with a C7g.xlarge instance costing $0.1360/hour on-demand versus $0.1700/hour for a comparable C6i.xlarge.</li> </ul> <p>The x86 era is not over, but it has lost the momentum. The shift that Apple started in 2020 with the M1 has completed its first arc: ARM-based silicon now dominates the high-performance laptop market, is the cost-performance leader in cloud compute, and is rapidly infiltrating the data center. If you write software professionally, ignoring this transition is no longer viable.</p> <h2> How We Got Here </h2> <p>Apple Silicon was the visible inflection point, but the pressures driving it had been building for a decade. x86 power efficiency had plateaued. Intel's process technology roadmap had repeatedly slipped. And ARM Holdings had been quietly improving the ISA and licensing it to increasingly sophisticated chip designers — Apple, AWS, Qualcomm, Microsoft, Ampere — who were outpacing Intel's in-house teams on performance-per-watt.</p> <p>The M1's performance metrics in late 2020 were not incremental. Single-core performance matched or exceeded Intel's fastest laptop chips at roughly half the power draw. The M1 Pro, M1 Max, M2, M3, and now M4 series have compounded that lead. The MacBook Pro is now the default machine for a large percentage of professional developers, not because of Apple loyalty, but because the hardware is objectively better for most workloads.</p> <h2> AWS Graviton: The Cloud Story </h2> <p>On the cloud side, AWS Graviton has been the most consequential development for production infrastructure. Graviton3 instances (C7g, M7g, R7g) launched in 2022 offered 25% better compute performance and up to 60% better energy efficiency than comparable x86 instances at a 10–20% lower price point. Graviton4, launched in 2024, pushed further — claiming up to 40% better performance than Graviton3 for compute-intensive workloads.</p> <p>The pricing differential matters for anyone running significant cloud spend. A C7g.xlarge (Graviton3) costs $0.1360/hour on-demand versus $0.1700/hour for a comparable C6i.xlarge (Intel Ice Lake) — roughly 20% cheaper before you factor in the performance per dollar. At scale, this is not a marginal difference. Organizations running multi-million dollar AWS bills have moved 50–80% of their fleet to Graviton and are reporting meaningful cost reductions.</p> <h2> Microsoft's Cobalt and the Hyperscaler Pattern </h2> <p>Microsoft's Cobalt 100 — their custom ARM-based chip for Azure — went into production in late 2023. Google has its Axion processor. Meta has started custom silicon work. The pattern is clear: every major hyperscaler is vertically integrating their compute at the chip level, and they are all doing it on ARM.</p> <p>This is not coincidence. The economics of custom ARM silicon for hyperscalers are compelling: ARM's licensing model gives them control over microarchitecture decisions that are specific to their workload mix, and ARM's power efficiency allows higher rack density. When your data center has hundreds of thousands of servers, shaving 15 watts per server translates to tens of millions of dollars in annual power costs.</p> <h2> What This Means for the Software You Write </h2> <h3> The Compatibility Problem (Mostly Solved) </h3> <p>The biggest concern two years ago — binary compatibility — is now largely a non-issue for most software stacks. The major language runtimes (Node.js, Python, Go, Java, Rust, .NET) all have ARM64 builds that are well-tested. Docker supports multi-arch images with buildx. GitHub Actions and most CI providers support ARM runners. If you are writing in a high-level language, your code almost certainly runs on ARM without modification.</p> <p>Where it still bites: native extensions, compiled C libraries embedded in Python packages, and any code with x86-specific SIMD intrinsics hardcoded. If you have a data science stack with obscure native dependencies, test it on ARM before you migrate — there are still rough edges.</p> <h3> Build Pipelines Need Updating </h3> <p>The most immediate practical change for most teams: Docker images built on x86 run on ARM via QEMU emulation, but emulated x86 on ARM is slow — sometimes 3–5x slower for compute-heavy operations. If you are building images that will deploy to Graviton or Apple Silicon, you need native ARM builds.</p> <p>The right move: add ARM64 to your docker buildx build targets and push multi-arch images to your registry. The GitHub Actions ARM runners (now generally available for public repos) make this straightforward to add to existing CI pipelines. A linux/amd64,linux/arm64 build matrix adds maybe 10 minutes to a typical CI pipeline and future-proofs your images.</p> <h3> The Development-Production Parity Win </h3> <p>If your team is on Apple Silicon (M-series Macs) and your production runs on Graviton, you have ARM-to-ARM parity in development and production for the first time. This eliminates an entire class of "works on my machine" bugs related to architecture differences. Native ARM development environments running against native ARM production is a meaningfully better situation than the x86 emulation or architecture mismatch that most teams dealt with from 2020–2023.</p> <h2> The Software You Run on Servers </h2> <p>For common infrastructure software, ARM support is now first-class: PostgreSQL, Redis, Nginx, Kafka, Elasticsearch — all have battle-tested ARM64 builds. The tail risk is specialized proprietary software or legacy enterprise tools where vendors have not prioritized ARM builds. If you run anything in that category, check vendor support status before a Graviton migration.</p> <p>The trajectory is clear. ARM is not a niche architecture anymore — it is the efficient frontier of compute for both client hardware and cloud infrastructure. The developers who understand what that means for their build pipelines, their dependency stacks, and their production deployments are going to be running faster, cheaper, and more reliably than the ones still thinking of ARM as "the Mac thing."</p> <h2> Related reading </h2> <ul> <li><a href="https://tech.crescevo.com/migrating-from-go-to-rust/" rel="noopener noreferrer">Migrating from Go to Rust</a></li> <li><a href="https://tech.crescevo.com/microsoft-s-6502-basic-is-now-open-sourc/" rel="noopener noreferrer">Microsoft's 6502 BASIC is now Open Source (2025)</a></li> <li><a href="https://tech.crescevo.com/california-moves-to-exempt-linux-from-it/" rel="noopener noreferrer">California moves to exempt Linux from its age-verification law after backlash</a></li> </ul> <h2> Before you go — a free checklist </h2> <p>Shipping a service? <a href="https://tech.crescevo.com/production-readiness-checklist/" rel="noopener noreferrer">The Production Readiness Checklist</a> covers the 30 things to verify before it goes live — reliability, observability, deploys, security, data, and on-call. Subscribe free to unlock it.</p> <p><em>Originally published on <a href="https://tech.crescevo.com/the-arm-takeover-is-complete-what-it-means-for-developers/" rel="noopener noreferrer">Tech at Crescevo</a> — subscribe free for more.</em></p> programming devops webdev tech ironbyte-rgb Mon, 22 Jun 2026 04:00:15 +0000 https://dev.to/crescevo/the-engineering-managers-guide-to-ai-code-review-2og5 https://dev.to/crescevo/the-engineering-managers-guide-to-ai-code-review-2og5 <h2> <strong>TL;DR</strong> </h2> <ul> <li>AI code review tools are most effective in tasks such as pattern detection at scale, documentation gaps, and security anti-patterns, including issues like hardcoded credentials and SQL injection vectors.</li> <li>The use of AI code review for architecture decisions often adds noise, as AI reviewers are trained on code and not on business context or team conventions, leading to suggestions that may be technically correct but contextually wrong.</li> <li>Most teams that initially deployed AI code review at full verbosity, resulting in a high volume of comments, rolled back to a filtered mode within 60 days to avoid burying legitimate issues in noise.</li> </ul> <p>The AI code review landscape in 2026 looks nothing like the hype from two years ago. The tools have matured, the use cases have clarified, and the teams that have gotten real value from them have mostly figured out what to use, what to skip, and what category of problem they should not be asking AI to solve.</p> <h2> What AI Code Review Actually Does Well </h2> <p>The highest-value use cases for AI in the code review loop are pattern detection at scale, documentation gaps, and security anti-patterns. These are tasks that human reviewers do inconsistently because they are tedious, not because they are difficult.</p> <p>Security anti-patterns specifically: hardcoded credentials, SQL injection vectors, unsafe deserialization, missing input validation. AI reviewers catch these reliably and do not get tired of looking for them on the 400th PR of the quarter. This is where the ROI is clearest and the noise-to-signal ratio is lowest.</p> <p>Documentation gaps are the second clear win. AI reviewers are good at flagging missing docstrings, undocumented function parameters, and public APIs with no usage examples. These are exactly the tasks that human reviewers deprioritize under deadline pressure, and exactly the tasks that create maintenance debt six months later.</p> <h2> Where AI Code Review Adds Noise </h2> <p>Architecture decisions. AI reviewers are trained on code, not on business context, team conventions, or the tradeoffs that led to the current system design. A suggestion to refactor a module for better separation of concerns may be technically correct and contextually wrong.</p> <p>AI suggestions on architecture-adjacent code produce comment volume that engineers learn to ignore. Once engineers start ignoring comments, they start ignoring all comments, including the legitimate security and quality issues. This is the most dangerous failure mode: not that AI review produces bad suggestions, but that it produces so many suggestions that the signal gets buried in noise.</p> <p>The practical rule: configure AI code review to surface findings above a severity threshold, not to comment on everything. Most teams that deployed AI code review at full verbosity rolled back to a filtered mode within 60 days because engineer fatigue with AI comments degraded overall review quality.</p> <h2> Tool-Specific Notes </h2> <p>GitHub Copilot Code Review is strongest on JavaScript, TypeScript, and Python where the training data is densest. Weakest on infrastructure code (Terraform, Helm charts) and domain-specific languages. Good for catching common patterns; misses context-dependent architectural issues.</p> <p>Claude via API in CI pipeline is most flexible. You define the review criteria via system prompt, which allows you to encode team conventions and project-specific rules that generic tools cannot know. Requires more setup but produces the highest signal-to-noise ratio when the system prompt is well-configured.</p> <p>Cursor is primarily a development tool, not a code review tool. Its review capabilities are IDE-based and designed for the author of the code, not for async review by a second developer.</p> <h2> The Management Decision </h2> <p>The question engineering managers should be asking is not which AI code review tool is best. It is: what review tasks are we currently doing poorly, and would AI do them more consistently than a tired human reviewer?</p> <p>If the answer is security patterns and documentation, deploy AI code review with a focused prompt and a high severity threshold. If the answer is architecture quality and system design, AI code review is not the tool. Better design review processes and RFC culture are the tool.</p> <h2> Related reading </h2> <ul> <li><a href="https://tech.crescevo.com/logs-to-observability-playbook/" rel="noopener noreferrer">From Raw Logs to Reliability: An Engineering Playbook for Observability</a></li> <li><a href="https://tech.crescevo.com/the-engineering-hiring-screen-that-predicts-actual-job-performance/" rel="noopener noreferrer">The Engineering Hiring Screen That Predicts Actual Job Performance</a></li> <li><a href="https://tech.crescevo.com/welcome-to-vel-tech/" rel="noopener noreferrer">Welcome to Engineering Weekly</a></li> </ul> <h2> Before you go — a free checklist </h2> <p>Shipping a service? <a href="https://tech.crescevo.com/production-readiness-checklist/" rel="noopener noreferrer">The Production Readiness Checklist</a> covers the 30 things to verify before it goes live — reliability, observability, deploys, security, data, and on-call. Subscribe free to unlock it.</p> <p><em>Originally published on <a href="https://tech.crescevo.com/engineering-manager-guide-ai-code-review/" rel="noopener noreferrer">Tech at Crescevo</a> — subscribe free for more.</em></p> programming devops webdev tech ironbyte-rgb Mon, 22 Jun 2026 03:00:15 +0000 https://dev.to/crescevo/why-your-monolith-is-not-the-problem-1mkc https://dev.to/crescevo/why-your-monolith-is-not-the-problem-1mkc <h2> <strong>TL;DR</strong> </h2> <ul> <li>Teams that decompose a monolith into microservices prematurely often end up with a distributed monolith, managing around 30 services, two years later.</li> <li>The common complaints about monoliths, such as slow build times, can be solved using techniques like incremental compilation, build caching with tools like Turborepo, Gradle, or Buck, and parallel test execution.</li> <li>A modular monolith pattern can be enforced using language-specific mechanisms, such as module visibility rules and JPMS in Java or Kotlin, or explicit package structure and import linting in Python.</li> </ul> <p>Microservices have been the dominant architectural recommendation in software engineering for a decade. The case for microservices, independent deployability, team autonomy, technology flexibility, is well-understood. Less well-understood is what actually happens when teams decompose a working monolith into microservices prematurely, and how many of the failures blamed on the monolith are actually failures of a different kind.</p> <p>The pattern is consistent enough to be diagnostic: a team has a monolith that is getting hard to work with. They read about microservices, get excited, and begin decomposition. Two years later, they have a distributed monolith, all the coupling of the original plus the operational complexity of managing 30 services plus distributed transaction problems that did not exist before.</p> <h2> What Monolith Problems Actually Look Like </h2> <p>Before decomposing anything, it is worth being precise about which monolith problems you actually have. The common complaints are: slow build times, long test suites, deployment coupling (one bug blocks all deployments), and team coupling (two teams constantly stepping on each other's code).</p> <p>Each of these has solutions that do not require service decomposition. Slow build times: incremental compilation, build caching (Turborepo, Gradle, Buck), and parallel test execution. Deployment coupling: feature flags and separate deployment tracks within the monolith. Team coupling: module-level ownership enforcement, documented interfaces between modules, and architecture fitness functions that prevent cross-module coupling from accreting silently.</p> <h2> The Modular Monolith Pattern </h2> <p>The modular monolith is the pattern that teams usually wish they had built before decomposing. It is a single deployable unit with strictly enforced module boundaries. Modules can only communicate through defined interfaces, not by direct function call or shared database table.</p> <p>The enforcement mechanism varies by language. In Java or Kotlin, module visibility rules and JPMS can enforce this. In TypeScript, barrel exports and ESLint rules that block cross-module imports. In Python, explicit package structure and import linting.</p> <p>A modular monolith solves the team coupling problem without introducing distributed systems complexity. It also creates the right foundation for future decomposition: when a module genuinely needs to be extracted into a service, the interface already exists and the extraction is straightforward.</p> <h2> The Strangler Fig: When Decomposition Makes Sense </h2> <p>The strangler fig pattern is the right decomposition approach for monoliths that genuinely need decomposition. The principle: extract functionality incrementally, routing requests to new services while the old monolith handles the rest. Never do a big-bang rewrite.</p> <p>The practical version: put a proxy layer in front of your monolith. Route specific request paths to new services as they are built. The monolith handles everything the proxy does not explicitly route elsewhere. Over time, the monolith shrinks and the new services expand. You can stop at any point and still have a working system.</p> <h2> When Microservices Actually Make Sense </h2> <p>The cases where microservices are genuinely warranted: dramatically different scaling requirements, genuinely separate team structures operating with different release cadences, and components that need to be developed in different languages because the problem domain favors a different runtime.</p> <p>Size is not a decomposition criterion. Coupling, scaling requirements, and team independence are decomposition criteria. A 500K line modular monolith with clean module boundaries and fast build times is a better engineering platform than 50 microservices with unclear ownership, cross-service transaction complexity, and a Kubernetes cluster that costs $3K per month to operate.</p> <h2> Related reading </h2> <ul> <li><a href="https://tech.crescevo.com/migrating-from-go-to-rust/" rel="noopener noreferrer">Migrating from Go to Rust</a></li> <li><a href="https://tech.crescevo.com/microsoft-s-6502-basic-is-now-open-sourc/" rel="noopener noreferrer">Microsoft's 6502 BASIC is now Open Source (2025)</a></li> <li><a href="https://tech.crescevo.com/california-moves-to-exempt-linux-from-it/" rel="noopener noreferrer">California moves to exempt Linux from its age-verification law after backlash</a></li> </ul> <h2> Before you go — a free checklist </h2> <p>Shipping a service? <a href="https://tech.crescevo.com/production-readiness-checklist/" rel="noopener noreferrer">The Production Readiness Checklist</a> covers the 30 things to verify before it goes live — reliability, observability, deploys, security, data, and on-call. Subscribe free to unlock it.</p> <p><em>Originally published on <a href="https://tech.crescevo.com/monolith-not-the-problem/" rel="noopener noreferrer">Tech at Crescevo</a> — subscribe free for more.</em></p> programming devops webdev tech