hot events · 2026-07-18

The interesting bit isn't the criticism itself—it's the setup. Sam Altman invited Eggers to speak internally, and Eggers used the slot to fire at the product. His argument: thinking happens in the messy, embarrassing first-draft stage, and ChatGPT replaces that with editing AI output, so people stop practicing. He's not talking about model capability or copyright; he's laser-focused on cognitive deprivation. The article doesn't say how OpenAI staff or Altman responded, which is a real gap—the reaction to being called out in your own office might be more revealing than the critique.

Max Woolf lays it out plainly: he pays $100/month for Codex, and OpenAI reset his weekly quota six times in two weeks—often when he still had 50%+ left, effectively wasting $12 each time. His read is that this isn't outage compensation; it's a deliberate move to refill your quota before you run out, so you never have a dry spell that might send you to Claude Code or a competitor. I buy this interpretation—the timing is too convenient, and the resets are only announced through individual employee accounts, not official channels. The post doesn't give Anthropic's reset count, so we can't compare, but Codex's frequency alone looks more like retention engineering than goodwill. Woolf is now considering downgrading from $100 to $20 because the higher quota keeps getting wiped before he can use it.

This one's worth opening because it's a near-replay of the DeepSeek R1 moment from early 2025: a Chinese company open-sources a model, independent benchmarks put it close to the top closed models, and US chip stocks dip. Moonshot AI released Kimi K3 this week, and Arena.ai and Vals AI scores show it trading blows with Claude Fable 5 and GPT-5.6 Sol—though the company itself says it still trails. I'd discount the panic a bit. The post doesn't disclose parameter count, training cost, or the open-source license. Those three numbers are what would tell us how real the threat is. Right now we have benchmark scores and a stock move, which reads more like the market looking for a narrative than a model suddenly crossing a decisive line.

This post lands because the author isn't running a benchmark—he just used both models for his normal coding work and noticed they're indistinguishable on output quality and token count. The price gap is stark: K3 API at $3/$15 per million input/output tokens vs Claude's $10/$50. Subscriptions are worse—Kimi's $39 coding tier is generous, while Claude's $20 plan quietly swapped Fable for Opus because the economics didn't work. He pulls the lens back to policy: US restrictions on American models only constrain American customers. Semgrep found GLM 5.2 beating Claude on cyber benchmarks precisely because the restricted model declines work the open one just does. His prediction: the US will replay its auto-industry playbook—subsidies and tariffs propping up domestic models that can't compete abroad. I'd discount the policy prediction a bit since it's one person's extrapolation, but the price data is public and the Fable downgrade on the $20 plan is a documented fact. Those two points alone are enough to make anyone paying for Claude coding work recalculate.

This post is worth clicking because the author ran Claude Fable 5 and GPT-5.6 Sol on an unpublished fiber-network optimization problem, three 30-minute trials each, plain vs /goal. Fable 5's plain mode averaged 32,386—1,875 points better than Sol's 34,261—and its three runs stayed within a 319-point spread. That kind of consistency on an NP-hard problem is rare. /goal won four of six head-to-heads but dragged both models' averages down: Fable 5 by 759 points, Sol by 868. The author also cracked open how /goal works under the hood: Claude Code uses Haiku as a transcript-only evaluator, while Codex has persisted state and lifecycle tools—same command, completely different machinery. I'd discount this a bit: one person, one unpublished problem, small sample. But the stability numbers on Fable 5 are genuinely striking, and the takeaway that /goal isn't a safe default holds up.

This isn't analysis—it's a sentiment sample from a working engineer. Samuel Sutch claims 100% of engineering output in many orgs is AI slop, and even leadership roadmaps are at least half AI-generated. The sharpest detail: companies now screen for AI enthusiasm in the first interview. Answer wrong and you're flagged as difficult. He points to Zig as a rare holdout but expects non-AI work to stay niche. I'd read this as a phenomenon, not a verdict. He doesn't define slop—is it bad code, or just code he didn't write? No data either. But he's right about one thing: AI has shifted from tool to compliance requirement. If you're not using it, you're the problem. For engineers who love the craft, that stings. If you're in a similar spot, this at least tells you you're not alone.

I'd open this one because it reframes DSLs away from the tired “should we build a language” debate and toward something more practical: giving an LLM a constrained environment where it can't wander off. The author built telescope control software, abandoned DSLs for Python sandboxing back then, and now argues AI has dropped the cost of designing languages, writing parsers, and preparing examples—so teams can build domain boundaries first and decide later whether to keep them. The stack is clean: FastAPI handles atomic actions, DSL describes complete plans, Context Infrastructure stores past judgments and corrections. Tickloom shows the mechanism works, but the post is honest that there's no cross-model benchmark proving DSLs improve accuracy. Where I'd discount a bit: the warning about boundaries being too narrow and excluding correct answers is right, but the post doesn't give concrete heuristics for judging boundary width—it's more a principle than a method. Worth reading as architecture thinking, not as a performance claim.