AI news

The Future of Everything is Lies, I Guess (aphyr.com) AI

The author argues that today’s AI—especially large language models—is less like human-like intelligence and more like a “bullshit machine” that statistically imitates text while frequently confabulating, misunderstanding context, and making factual errors. They describe why LLMs can’t reliably reason about their own outputs, how “reasoning traces” and generated explanations can be misleading, and why models can be both astonishingly capable and still repeatedly “idiotic” in practical tasks. Overall, the piece frames current AI progress as creating major real-world risks alongside potential benefits, without offering a single definitive prediction of the future.

LLM plays an 8-bit Commander X16 game using structured "smart senses" (pvp-ai.russell-harper.com) AI

Russell Harper describes bringing his 1990 8-bit PvP-AI game to the Commander X16 (first running well in an emulator but slower on real hardware due to a rendering issue). He also explains how an LLM can play the game using structured “smart senses” instead of raw visual input, with the game converted to turn-based and equipped with text-based state inputs. The write-up covers the PHP-to-emulator integration and reports a progression in LLM strategies across recorded games.

MegaTrain: Full Precision Training of 100B+ Parameter LLMs on a Single GPU (arxiv.org) AI

MegaTrain is a proposed training system that enables full-precision training of 100B+ parameter LLMs on a single GPU by keeping model parameters and optimizer states in CPU host memory and streaming them layer-by-layer to the GPU for computation. The method uses double-buffered pipelining to overlap parameter prefetching, gradient computation, and offloading, and it avoids persistent autograd graphs via stateless layer templates. Reported results include training up to 120B parameters on an NVIDIA H200 with 1.5TB of host memory, and improved throughput versus DeepSpeed ZeRO-3 with CPU offloading on smaller models.

Mario and Earendil (lucumr.pocoo.org) AI

Armin Ronacher announces that Mario Zechner is joining the Earendil team, praising Pi as a thoughtful, quality-focused agent infrastructure and contrasting it with the industry’s rush for speed. He links the hire to concerns about AI systems producing “slop” and degradation, and describes Earendil’s Lefos effort to build more deliberate tools that improve communication and human relationships. Ronacher says he and Colin want to steward Pi as high-quality, open, extensible software while clarifying how it may relate to Lefos.

Multi-agentic Software Development is a Distributed Systems Problem (AGI can't save you) (kirancodes.me) AI

The post argues that multi-agent software development with LLMs is fundamentally a distributed systems coordination problem, not something that “smarter agents” will eliminate. It models prompt-driven code synthesis and agent collaboration as an underlying consensus task constrained by an underspecified natural-language spec, then relates the setting to classic impossibility results like FLP (showing limits on deterministic consensus under async delays and possible crashes) and discusses possible parallels to failure detectors. The author concludes that building scalable tooling/languages for agent coordination remains necessary even if future models become extremely capable.

The Downfall and Enshittification of Microsoft in 2026 (caio.ca) AI

The article argues that Microsoft’s 2026 “enshittification” is driven by shifting focus from core product quality to aggressive, AI-centered Copilot integration across Windows, Office, and GitHub. It points to Windows 11 promises to fix long-standing desktop usability issues, recurring complaints and outages affecting GitHub’s developer workflows, and the perceived tradeoff between reliability and Copilot placement. The author also suggests competitive pressure from Apple’s lower-cost MacBook Neo and Linux’s gradual desktop legitimacy is making Microsoft’s strategy look less like leadership and more like defensive retrenchment.

I've Sold Out (mariozechner.at) AI

Mario Zechner says he has joined the Earendil team and will “take pi” as a coding agent, explaining his history of OSS-to-commercial transitions and the pain he saw when key projects like RoboVM went closed-source after being sold. He describes growing interest from VCs and large companies in pi, but says he does not want to run a VC-funded company focused only on pi, prioritizing family time and avoiding the stress and community-betraying dynamics he experienced before. The post also recounts how Zechner met Armin and others in the “Vienna School of Agentic Coding” circle and how collaboration around agentic coding led to this decision.

Open Models have crossed a threshold (blog.langchain.com) AI

LangChain reports early Deep Agents evaluations showing open-weight models such as GLM-5 and MiniMax M2.7 can match closed frontier models on core agent abilities like file operations, tool use, and instruction following. The post emphasizes lower cost and latency, and describes how their shared eval suite and Deep Agents harness let developers compare and swap models across providers with minimal code changes.

An Arctic Road Trip Brings Vital Underground Networks into View (quantamagazine.org) AI

A Quanta Magazine field report follows biologist Michael Van Nuland and colleagues as they sample Alaskan tundra to test machine-learning predictions about rare mycorrhizal fungal “hot spots.” The article describes how underground fungal networks connect to plant roots, exchanging nutrients and carbon, and how recent imaging and robotic tracking suggest the fungi actively regulate this system rather than merely serving plants. Because these networks help store vast amounts of carbon in permafrost but are vulnerable to warming, wildfires, and thaw, the researchers argue that better mapping and protection of soil biodiversity could matter for climate resilience.

Japan relaxes privacy laws to make itself the 'easiest country to develop AI' (theregister.com) AI

Japan has approved amendments to its Personal Information Protection Act to remove the usual opt-in consent requirement for organizations using low-risk personal data for statistics and research, aiming to speed AI development. The changes include provisions for some sensitive categories such as health-related data (for improving public health) and facial images, with additional conditions like parental approval for children under 16 and stricter requirements around handling facial data. The rules add penalties for fraudulent data acquisition and improper use, but reduce requirements to notify individuals after data leaks deemed unlikely to cause harm.

Sonnet 4.6 Elevated Rate of Errors (status.claude.com) AI

Claude Status reports that Claude Sonnet 4.6 has an elevated rate of errors, affecting claude.ai, platform.claude.com, the Claude API, Claude Code, and Claude Cowork. The company says it is investigating the issue as of April 8, 2026, with incident updates available via email or SMS.

The BSDs in the AI Age (lists.nycbug.org) AI

The post proposes an NYC*BUG summer presentation and discussion thread on how AI and LLM tools are affecting work and security practices, including their impact on BSD operating systems and developers. It asks contributors about current LLM usage for everyday productivity, whether BSD projects should adopt explicit LLM-related policies (citing NetBSD’s commit guidance and credential-related CVE concerns), and how BSD teams and individuals might use LLMs for tasks like code discovery or vulnerability research.

Show HN: Can an AI model fit on a single pixel? (github.com) AI

Show HN shares an open-source project, ai-pixel, that trains a tiny single-neuron binary classifier and then encodes its learned weights into the RGB values of a downloadable 1x1 PNG. The demo lets users place training points, run gradient descent, and later load the “pixel model” to make predictions. The article emphasizes it’s an educational compression experiment with predictable limits (e.g., it can’t learn XOR or other non-linearly separable patterns).

Claude Is Dead (javiertordable.com) AI

The article argues that Anthropic’s Claude Code has been “nerfed” through cost-cutting changes—leading to faster rate-limit/token drain and reduced reliability for complex coding—prompting developers to complain publicly and switch to other tools or local models.

Hallucinated citations are polluting the scientific literature (nature.com) AI

Nature reports that large language models are increasingly generating fabricated or untraceable “hallucinated” references that have appeared in thousands of 2025 papers. An analysis of more than 4,000 publications found that many had invalid citations, and manual checks confirmed that 65 of the most suspicious papers contained at least one reference that could not be verified. The article also describes publisher screening efforts and the difficulty of deciding how to handle problems once such citations make it into the published record.

LLM scraper bots are overloading acme.com's HTTPS server (acme.com) AI

After intermittent outages in February–March, the ACME Updates author traced the issue to HTTPS traffic being overwhelmed by LLM scraper bots requesting many non-existent pages. When they temporarily closed port 443, the outages stopped, suggesting the slow HTTPS server and downstream congestion/NAT saturation were contributing. The author notes the same bot behavior is affecting other hobbyist sites and says a longer-term fix is needed.

New York Times Got Played by a Telehealth Scam and Called It the Future of AI (techdirt.com) AI

The article argues that a recent New York Times profile of Medvi, an “AI-powered” telehealth startup, relied on misleading framing—such as treating a projected revenue run-rate as a “$1.8 billion” valuation—while failing to report serious red flags. It claims Medvi’s marketing used deceptive tactics including AI-generated or deepfaked images and false credibility signals, and it notes regulatory scrutiny, including an FDA warning letter, plus lawsuits involving the company and partners. The author concludes the Times story elevated a narrative of AI-enabled entrepreneurship that doesn’t hold up under basic verification.

OpenAI says its new model GPT-2 is too dangerous to release (2019) (slate.com) AI

Slate reports that OpenAI withheld the full GPT-2 text-generation model, citing safety and security risks such as spam, impersonation, and fake news, while releasing only a smaller version. The article profiles GPT-2’s apparent capabilities and reviews expert skepticism that the danger may be overstated or that an embargo can meaningfully slow dissemination. It uses the controversy to highlight a broader debate over how to balance beneficial research and applications against the potential for misuse.

Ralph for Beginners (blog.engora.com) AI

The Engora Data Blog post explains how “Ralph” automates code generation by breaking a project into small, testable requirements from a product requirements document, regenerating code until each requirement’s acceptance criteria passes. It walks through setup (installing a codegen CLI, obtaining an LLM “skills” file, using git), converting a Markdown PRD into a JSON requirement list, and running a loop script that applies changes to the codebase and records pass/fail status without human intervention. The author cautions that results depend heavily on how thorough the up-front PRD is and notes that API costs and some rough setup/reporting still make experimentation nontrivial.

GLM-5.1 matches Opus 4.6 in agentic performance, at ~1/3 actual cost (app.uniclaw.ai) AI

The article reports benchmark results from OpenClaw Arena showing that GLM-5.1 achieves similar agentic performance to Opus 4.6 while using about one-third of the actual cost, highlighting a better cost-performance tradeoff. It also notes that some “provisional” models may have limited battle data and changing statistics as more results come in.

Larger and more instructable language models become less reliable (pmc.ncbi.nlm.nih.gov) AI

The article reports that as large language models have been scaled up and “shaped” with instruction tuning and human feedback, they have become less reliably aligned with human expectations. In particular, models increasingly produce plausible-sounding but wrong answers, including on difficult questions that human supervisors may miss, even though the models show improved stability to minor rephrasings. The authors argue that AI design needs a stronger focus on predictable error behavior, especially for high-stakes use.

We need re-learn what AI agent development tools are in 2026 (blog.n8n.io) AI

The article argues that by 2026 many core “AI agent builder” capabilities—like document grounding, evaluations integrations, and built-in web/file/tool features—have become table stakes via mainstream LLM products. It proposes updating agent development evaluation frameworks to focus more on enterprise-readiness (security, observability, access controls, sandboxing, reliability) and on how agents can operate deterministically within controlled workflows while still allowing safe autonomy like spawning sub-agents. The author also notes shifting emphasis away from MCP-style interoperability after security concerns, and suggests reassessing how coding agents should be evaluated versus their role inside broader automation pipelines.

AI Assistance Reduces Persistence and Hurts Independent Performance (arxiv.org) AI

A paper on arXiv reports results from randomized trials (N=1,222) showing that brief AI help can reduce people’s persistence and impair how well they perform when working without assistance. Across tasks like math reasoning and reading comprehension, participants who used AI performed better in the short term but were more likely to give up and did worse afterward without the system. The authors argue that expecting immediate answers from AI may limit the experience of working through difficulty, suggesting AI design should emphasize long-term learning scaffolds, not just instant responses.

What we learned about TEE security from auditing WhatsApp's Private Inference (blog.trailofbits.com) AI

Trail of Bits reports findings from an audit of Meta’s WhatsApp “Private Inference,” which uses TEEs to run AI message summarization without exposing plaintext to Meta. The review found 28 issues, including high-severity problems that could undermine the privacy model, and describes fixes focused on correctly measuring and validating inputs, verifying firmware patch levels, and ensuring attestations can’t be replayed. The authors argue TEEs can support privacy-preserving AI features, but security depends on many deployment details—such as input validation, attestation freshness, and negative testing—not just the underlying TEE isolation.

Show HN: Gemma 4 Multimodal Fine-Tuner for Apple Silicon (github.com) AI

The GitHub project “gemma-tuner-multimodal” describes a PyTorch/LoRA fine-tuning toolkit for Gemma 4 and Gemma 3n that targets multimodal data (text, images, and audio) on Apple Silicon using MPS/Metal, without requiring NVIDIA GPUs. It supports local CSV-based training (with streaming from cloud stores mentioned as an option) and exports fine-tuned adapters for use with HF/SafeTensors and related inference tooling. The repo also includes a CLI “wizard” for configuring datasets and launching training, plus installation guidance including a separate dependency path for Gemma 4.