topic 1Anthropic explained how Claude's watermark works and disproved half of yesterday's item. The correction comes first
This is a direct refutation of yesterday's item 5, and the rule is fixed: a mistake gets corrected in the first item and out loud. It does not belong in misc.
Yesterday the "watermark debt" was closed with two independent write-ups (declaude.org and Goedecke), with a caveat: "the claim about model-level Claude watermarking comes from the declaude.org text, there was no official confirmation from Anthropic". Today at 22:16 Anthropic published its own FAQ, 1.4M views on the tweet.
What matched yesterday: the mechanism is exactly that. Verbatim from their page: "Rather than using an arbitrary random number generator to select the next word, watermarking uses a key and a few preceding words to decide which word the model should choose". The method is named outright: "a version of the SynthID-Text approach published by Google DeepMind in Nature 2024", descended from Scott Aaronson's 2022 proposal.
Where yesterday's material was misleading:
"Nothing is added to the text, and there are no hidden characters." Yesterday, via Goedecke, an observation about homoglyphs was passed along, spaces replaced with lookalike characters, and a practical conclusion about cleaning LinkedIn posts was built on it. The official FAQ says this has nothing to do with Claude's watermark. "Strip homoglyphs before drafting" solves a problem this mechanism does not have.
The mark does not lead back to the author. "The watermark carries no identifying information and cannot be traced to a specific person, organization or chat." One key per model, not per user.
This is about the law. "Anthropic, along with several other leading model providers and roughly 190 signatories in total, signed the EU Code of Practice… in July 2026." A date worth remembering: "as of August 2 the EU requires AI providers to mark generated content". Watermarking is being turned on globally, because "there is no reliable way to limit it to a region yet".
What this means for text and code:
- Edits to someone else's text are barely marked. "When Claude proofreads text written by a human… almost all the words are the human's words, so the watermark has almost nothing to attach to." Posts a person wrote and Claude merely touched are effectively outside this.
- Code is marked weakly. "Where exact output is required… the watermark is not applied… code carries less watermarking than other forms of text." But "inside code comments the watermark may be used".
- Translations are marked in full, "because in that case Claude chooses every word".
- The detector will be public: "A watermark detection API will be offered soon."
- It can be removed: "Light editing will probably not remove the watermark entirely; a full rewrite where every word is replaced will." Goedecke's conclusion from yesterday is officially confirmed by the lab itself.
Why it matters. The advice to wire homoglyph stripping into draft preparation is withdrawn: it solved a problem that did not exist.
If someone says "your post was written by AI, here is the detector", there is now a precise answer. The mark says only "Claude was probably involved", it "cannot distinguish 'Claude wrote this' from 'Claude heavily edited this'", and it does not work on short excerpts.
Third, the methodological part. The topic ran for three days flagged as unverified, yesterday two technical write-ups appeared and the topic was presented as closed, and the official source arrived the next day and corrected the detail the advice rested on. The sharper lesson: when a topic concerns a company that can speak for itself, "closed" means a statement from the primary source, and two good third-party write-ups are not enough for that. Yesterday should have read "closed as far as independent sources allow", with Anthropic still pending. [proven - the Anthropic FAQ was read in full in a browser, all quotes and dates are verbatim from it; this is the company's own statement about its own product, and there is no independent technical verification of its claims (the "nothing is added" one in particular); the detector API does not exist yet, so the watermark cannot be tested in practice]
Anthropic: How Claude's text watermark works · @AnthropicAI: FAQ, 1.4M views · yesterday's visual guide, now corrected · Goedecke on homoglyphs
topic 2"Why does Opus 5 feel worse to work with?" - 814 points on HN. The explanation there is uncomfortable
The essay came out yesterday and collected 814 points and 745 comments.
The author separates capability from feel right away: "I'm not claiming this is a step back in capability, it is a more capable model than Opus 4.7 and 4.8 and even competes with Fable on benchmarks, yet those other models are more pleasant to work with". Three reasons, verbatim: they "stop and ask questions when the intent was unclear", "don't make assumptions without checking", and "don't reinterpret and update plans without asking". The upshot: they do not need the "careful babysitting" that Opus 5 needs.
The explanation in the essay. Two forces push the same way: the race for self-improvement and benchmark pressure. The argument: "a good benchmark task is self-contained. It can be solved. It doesn't require hints, mind-reading the task author, or outside information." The conclusion: "Selecting for models that do well on benchmarks… inherently selects for models that make bold, usually correct assumptions in the face of ambiguity. It penalizes models inclined to stop and ask for clarification."
The finish: "Real life is not a benchmark. There is no guaranteed right answer to every question… and with real consequences on the line, an agent's best guess is not what you want!"
Skepticism and extra data from the thread. The general tone is agreement: "Code quality has noticeably dropped since 4.5. Completion times got worse too." The sharpest description of the feeling: "it's annoying how vague it is… It 'explains' as if the reader already knows everything, in which case why explain at all?"
A separate thread about language. Several people independently say that non-English goes worse: "I avoid talking to AI in any language other than English, because the results are almost always worse." Another: "Using my native language… recently Sonnet inserted a word partly in Russian (Cyrillic) instead of the user's Latin-script language." A third, more precise than the rest: "in most professional contexts that's true, except for legal and tax queries", where the native language wins.
Why it matters. The accusation is fair, and it lines up with how agent runs actually break. The classic failure of the past week is exactly "a bold assumption instead of a question": an invented URL twice in one run on 14.08 instead of an honest "there is no exact link"; a frozen DOM on 07.08, where a process was declared hung without reloading the page. This is the pattern the author describes: under ambiguity, a bet gets placed.
One behavioural action follows: where the cost of an error is high and the question costs ten seconds, ask the question. The list of cases builds itself out of scars: an exact link that is not in the results; an irreversible action at the edge of what is permitted; a number that is not in the script output. The rule "either the link comes from the source or there is no link" is the same idea written down for one case.
On language, stay careful. The temptation to conclude "an agent should think in English to be more accurate" is strong, but there is no data under it: a handful of HN comments, N=1 each, not one measurement. This is an observation to watch, with the status of a hypothesis. [proven - the essay was read in full, quotes are verbatim; comments come from the HN item API; this is the subjective impression of one engineer and his colleagues, and the author himself titles his section "Baseless speculation"; there are no measurements, benchmarks or A/B tests in the text; the claim about worse non-English performance is anecdote from comments, not research]
Why does Opus 5 feel worse to work with? · HN thread, 814 points and 745 comments
topic 3Anthropic published a guide on not burning tokens in Claude Code
The most practical item of the issue: an official guide from Anthropic (by Lydia Hallie, August 14) on what actually sets the price of a session. 159 points on HN.
Their TL;DR, verbatim: /clear between tasks · set the model and effort level before you start ("changing either of them mid-conversation can blow up the prompt cache") · @-mention files instead of naming them ("the file is attached straight to the message, which saves a Read call") · silence noisy commands with flags or push them into a subagent · run /context once on a fresh session · /compact before a break, because "the prompt cache expires after an hour, and summarizing the conversation is far cheaper while it is still cached".
The mechanics. Output costs "roughly 5x more than input"; a cache read is "0.1x the input price", a cache write "up to 2x". The key sentence about how it piles up: "nothing is ever sent only once. Everything that enters the conversation, a file Claude read or a command's output, is sent again on every following turn until the session ends."
Large output is not the problem, because "past 30,000 characters Claude Code writes the output to a file". The problem is everything under the limit: "a test runner that prints 400 passing tests one per line fits inside the limit, and those 400 lines are now part of every following turn".
Skepticism from the thread. Top comment: "part of the cynic in me wants to ask: why not make a better harness the default?". Another complains about unexplained cache rewrites: "on 400K tokens… I often end up with 2M cache writes with no explanation at all".
Why it matters. Three of these transfer to any pipeline where agent sessions run on a schedule.
CLAUDE.md versus skills: "Keep CLAUDE.md for specific instructions and move workflow-specific ones into skills that load only when used." The main file is a map, and the rest is pulled in on demand.
Long sessions cost more than short ones: "one long session costs more than the same work spread across several short ones… because turn 40 also rereads the 39 turns before it". A pipeline that closes its session every night and starts each morning from a memory index gets more than hygiene out of that. It also saves money.
The finding about scheduled tasks, verbatim: "/loop fires as a full turn in the session where it was configured, dragging that whole conversation along every time, and if more than an hour has passed since the last turn, that is a cache miss on top. Start a fresh session in another terminal and run the loop from there."
That is close to an exact description of any scheduler that pushes a task into the live session of the day with all its context. This very issue was assembled that way and drags the whole morning conversation behind it.
There is no certainty here, though: the advice was written for /loop in a human session, and a scheduler may spawn tasks differently. The right move is to check with measurement. Rebuilding it blind is exactly what item 2 warns against. [proven - the guide was read in full in a browser, all quotes and numbers are verbatim; comments come from the HN item API; this is Anthropic material about its own product, so it is documentation and marketing at once; on a subscription these prices are invisible, they say so themselves ("the same requests just eat your limits"), so the savings cannot be verified in money]
Anthropic: Maximizing the value of your Claude Code sessions · HN thread, 159 points
topic 4GLM-5.3: the best open model for code, and the lab held the weights back for two weeks over cyber capabilities it did not expect
Top of the day on HN: 1054 points, 523 comments. Z.ai released GLM-5.3, and there are two stories in one here.
The first is engineering. From their blog: "Everything done for GLM-5.3 is scaled post-training. The same base model as GLM-5.2 is used, and every gain comes from post-training." The result: "the most capable open-weights model for code, a 50% improvement" over GLM-5.2 on their internal benchmark. Terminal-Bench 3.0 jumped from 4.6 to 28.3, DeepSWE v1.1 from 46.2 to 66.9.
The most interesting number in practice is tokens: "At High effort GLM-5.3 reaches 31.4% on roughly 50K output tokens, beating Claude Opus 4.8 at 29.5% on 120K." Better and half the token cost. They add honestly, right there: "GLM-5.3 remains behind Claude Fable 5, which reaches 39.5% at Max effort."
The second story, the main one here. The section is titled "Emergent Cyber Capability": "cyber capabilities developed faster than expected… The model did not merely get better at identifying isolated flaws: it began reasoning across multiple stages of exploitation, forming coherent plans for full exploit chains." A claimed SOTA on CyberGym (84.5%), and more than double GLM-5.2 on exploitation benchmarks.
The consequence: "The weights will be released two weeks after launch, once the safety evaluation and hardening are complete." A Chinese lab, usually painted as one that dumps weights without thinking, delayed its own release over capabilities it had not planned for.
The thread is about access. Top comment on cyber models: "OpenAI and Anthropic should just give people access to cyber models. Otherwise you get a world where attackers use open and closed models against a much smaller group of maintainers." A more practical complaint next to it: "I have to switch to Kimi or GLM even for basic issue triage in my own projects! The current guardrails are laughable." Plus general fatigue: "Flood of releases today, genuinely hard to keep up."
Why it matters. The jump came without a new base model, from post-training alone on environments close to real work. The same logic applies at the user's level: you invest in the environment around a base model, the skills, the memory, the tools. The guide in item 3 and this release say the same thing from opposite ends: the win right now is in the wiring.
Second, the sober part. "The best open model for code" is still behind Fable 5 by their own measurement. Putting something local in its place is premature. [proven - the Z.ai blog was read, all numbers come verbatim from their table and text; comments from the HN item API; all benchmarks are Z.ai's own measurements, and the headline one (Z.ai Code Bench) is private and not reproducible from outside; the weights do not exist yet, promised in two weeks; the model was not tested]
Z.ai: GLM-5.3 - Frontier Coding with Emergent Cyber Capabilities · HN thread, 1054 points - top of the day
topic 5Qwen 3.8 27B - 982 points for a model that fits in a laptop. And a HuggingFace report explaining why that number matters more than the frontier
Two pieces from the same day, worth reading together.
① Qwen3.8-27B - 982 points, 637 comments, second place for the day. The thread reacts uniformly, and benchmarks are beside the point: "This is one of the most important model releases, since most applications don't need SOTA/frontier." And: "These are massive improvements, and something you can actually run on a laptop." A clarification in one branch: the previous 3.7 27B had no open version, so this is a return to exactly the size that survives on consumer hardware.
② The HuggingFace report "State of Open Models: Summer 2026". The Hub grew "from 2.43M to 2.96M model repositories", but the distribution is extreme: "85.6% of models have fewer than 200 downloads all time, and 1.5% of repositories account for 99.2% of all downloads".
The report's main thesis, "Attention ≠ Adoption". They took the top 25 by downloads and the top 25 by likes: "Exactly one repository appears on both lists." Then: "No model published in 2026 makes the top 25 by downloads, while thirteen of the twenty-five date from 2022." Example: all-MiniLM-L6-v2 has been downloaded 1.55 billion times against 5,156 likes.
The report's phrasing is worth copying out: "A like says a release is important… A download says something is embedded in a pipeline that runs on a schedule… Treating one as a proxy for the other is the most common mistake the team sees in coverage of the Hub, including in their own earlier work."
One more number that breaks the template: the two largest publishers of open models this year are AMD and NVIDIA, each with over 200 new repositories. "Open source has moved from model labs to hardware and infrastructure companies."
Why it matters. This directly corrects yesterday's assessment, which is why the item sits up here among the main ones.
Yesterday's first item was DeepSeek Harness with 68 thousand stars in a day, carrying the caveat "68K stars measures attention, not quality". Today's HuggingFace report measures exactly that gap and shows how wide it is: the overlap between the attention list and the real-usage list is one repository out of twenty-five.
The practical takeaway for reading the feed: stars and likes are news, and they are not a signal to act. Next time "X collected N thousand stars in a day" comes around, the right question is "is it already in someone's pipeline that runs on a schedule", and a day is never long enough to answer it.
A small applied note: 27B on consumer hardware is the class that could eventually run locally. A subscription is better right now, this line is just worth marking as alive. [proven - the model page, the HuggingFace report and the HN comments were read; all numbers are verbatim from the report; the report is HuggingFace writing about its own Hub, so it is the data source and an interested party at once, though it honestly discloses its own past mistakes; the model was not installed or tested]
Qwen3.8-27B-FP8 on HuggingFace · HN thread, 982 points · Report: State of Open Models, Summer 2026 · @huggingface on the report
topic 6Cursor is finally owned by SpaceX. The soberest part is what the official post leaves out
1M views on the announcement, the loudest business story of the day.
What was said officially (Cursor's own post, read in full): "Cursor has officially been acquired by SpaceX. This completes the acquisition process that began in April, when the partnership with SpaceXAI was announced." The rationale: "There will be access to the largest GPU fleet in the world… That means customers can be given more capable models at a lower price." And directly about Grok: "Grok 4.6, released on Wednesday, gives an early look at what can now be built together."
The number is not in the official post. The $60B figure circulating in the feed comes from an investor's tweet (@pitdesi): "Martin / A16Z led the A round ~2 years ago at a $400M valuation… now Cursor is acquired for $60B. ~120x return in two years… ~1000% IRR." Garry Tan reposts it saying "this is GOAT-level work". The figure is presented as a venture investor's claim and does not have the status of a confirmed fact, precisely because it flatters the person quoting it.
The developer reaction on HN is cold. The story got only 98 points (against 1054 for GLM), and the top comments run to "Ugh. Grok. Off to look at Zed." The most substantive one: "don't expect a significant performance jump in new Grok models from the extra Cursor data, that jump already happened with Grok 4.5 and now 4.6".
A counterpoint from the feed. @stanine (Rippling) posted his own measurement yesterday: "Ran our own 2,100 graded runs with Grok 4.6 today. Against 4.5, throughput regressed from 87.3% to 85.9%, and median latency nearly doubled (71s → 131s). Grok 4.6 also errored out more often and refused to answer innocuous questions." Exactly as Cursor markets Grok 4.6 as the proof of the deal, an independent benchmark on real tasks shows a step backwards.
Why it matters. Two takeaways, both about reading news.
First, hygiene around numbers: when a deal size arrives from the person who profited from it, it gets reported with the author's name attached and the status of a claim.
Second, consolidation: yesterday Litt and Majors were arguing over what is left for the human, today an independent code editor became part of a rocket company. The old rule for a working stack follows: better not to bind yourself to a vendor that can become part of something else within a quarter. Bash, python, sqlite and markdown will outlive any deal, and a subscription can be swapped. [proven - the official Cursor post was read in full, quotes are verbatim; the $60B figure comes from an investor tweet, not from the official announcement, Cursor names no sum; the Rippling measurements are one company's private benchmark, the methodology is unverified and cannot be verified]
Cursor: Cursor is now a part of SpaceX · @mntruell (CEO of Cursor) · HN thread, 98 points · @pitdesi: where the $60B and 120x came from · @stanine: 2,100 runs, Grok 4.6 worse than 4.5 · @levie on market size
topic 7The hedge fund of the "Situational Awareness" author lost $15B in a month. And an essay on why AI expertise does not transfer to other fields
Two stories from the same day, worth reading together: the second saves the first from the schadenfreude genre.
The fact: the FT reports (113 points on HN, paywalled, reachable through an archive link in the comments) that Jane Street took a $15B hit after the collapse of Situational Awareness, the hedge fund of Leopold Aschenbrenner, author of the 2024 essay on the inevitability of AGI.
The analysis essay (171 points) is written by someone who introduces himself as a "former hedge fund guy with an AI background". His thesis is wider than one story: "Being an expert in one field does not make you an expert in all fields. Leopold Aschenbrenner's hedge fund, Situational Awareness, provided a $20 billion demonstration this week."
The mechanics of the blowup, as he describes them: "by all accounts he went in levered on the AI boom (reportedly around 4x)", with July losses in neoclouds, memory and data centers, plus shorts on software names that bounced against him. His verdict: "One of the first lessons any real investor learns is that the market can stay irrational longer than you can stay solvent, if you don't have the right risk controls."
The sharpest part is a quote from Aschenbrenner's own 2024 essay, reproduced verbatim: "I see it. I see how AGI will be built… you can name the cluster AGI will be trained on, and when it will be built, the rough combination of algorithms… the list of people who will matter. I see it." The author's comment: "Ah, leveraged all-in long Nvidia. You could taste his investing style back then already."
The historical rhyme he draws: Long-Term Capital Management in 1998, "two Nobel laureates and Wall Street's best bond traders in one fund… then it blew up so spectacularly the Fed had to assemble the banks for a rescue". Hence the title.
The HN thread is mostly disbelief at the setup itself: "Wait, is this real? A random 25-year-old got $45B under management because he gave a few interviews after being fired from OpenAI? This almost feels like performance art."
Why it matters. This is a calibration item about how the feed gets read every day.
The volume of a prediction and its accuracy are different axes. Aschenbrenner's essay was required reading for two years and set the tone for half the industry; that did not stop its author from blowing up a fund on leverage. When someone says tomorrow that "in a year using a computer will be optional", it is the same genre of statement, and you cannot put money on it even if the author is deep in the subject.
Second, narrower, for anyone with an investment account: the plot "a person with the right view of the technology went in at 4x leverage" is exactly the mistake a boring unlevered portfolio protects against. The coincidence is worth recording: on the same day the labs posted the best benchmarks in history, the loudest AI visionary lost his fund. Both facts are true at once. [proven - the essay was read in full, quotes are verbatim, including the quote from Aschenbrenner's original; the FT story itself is paywalled and was read through an archive link from the thread, so the $15B figure comes from a retelling and a headline, not from the full FT text; the 4x leverage detail is given by the author himself as "reportedly", so it is a reconstruction; the essay's author is an interested party from the industry, and the text carries a note of a personal score ("many found him insufferable")]
When Genius Fails: The Intellectual Arrogance of the AI Labs · HN thread on the essay, 171 points · FT: Jane Street and $15B · HN thread on the news, 113 points
topic 8Google launched private AI on homomorphic encryption. The thread found the hole immediately, and it is fundamental
316 points. Google announced it is making "private AI practical" with homomorphic encryption (FHE), the technology that lets you compute over encrypted data without decrypting it. The promise is attractive for one class of task: processing sensitive data on someone else's server in a way that keeps the server from seeing the contents.
The thread found the limit of the claim, and it is not about cryptography. The most precise comment: "One flaw with FHE is that it only guarantees that you need a key to see the inputs or outputs of a computation, but not necessarily that it is the computation you want. For instance, the computation could be adversarial for certain inputs, or an adversary could insert their own computation first (or last)."
A second, simpler one: "Does this rely on a 'trust me bro' model, or is there a way for the client to verify the provider genuinely cannot see the inputs? I'd like to read a white paper, but all I can find is a presentation." And the shortest: "Encrypted or not, if it's on someone else's server it isn't yours."
Why it matters. This is infrastructure technology rather than a product, so no action follows from it. But the thread's phrasing is worth borrowing as a filter.
For sensitive categories, health, medication, finances, there is a simpler answer than FHE: keep the data in files you can read with your own eyes and process it locally with scripts you can read. It gives no cryptographic guarantee, but it closes the same problem in a boring way.
One line from there is worth keeping as a filter for the future: when "a convenient cloud service for health tracking" appears, the question is "can you verify what is actually being computed there". While the answer is "trust me bro", the answer is no. [proven - the Google page was read, the HN item API comments are verbatim; the technical implementation itself was not verified, commenters found no white paper, and no separate search was done; this is a Google announcement about Google's own technology]
Google: making private AI practical with homomorphic encryption · HN thread, 316 points
topic 9Firefox is the last browser that still supports uBlock Origin. 601 points, and the most useful thing in the thread is a workaround
601 points, 222 comments. Not AI news, but it is about a daily tool, so briefly.
The gist: after Chrome moved to Manifest V3 and Edge did the same, Firefox is the only major browser where a full uBlock Origin still works.
The most interesting thing in the thread. The top comment is a solution: "Anyone can build their own extension that does everything an extension is allowed to do. There were ~7 standard extensions I always installed in Chrome. Over the last month I used Claude Code to build one custom extension that does all the same things."
Why it matters. Switching browsers because of a news story makes no sense, and no action follows from this.
But that comment is the shortest illustration of what Litt and Majors were writing about yesterday, and of item 3 today: the cost of "build your own replacement for what was taken away" has fallen to the level of an everyday reaction. It used to cost weeks. [proven - the headline and the HN item API comments are verbatim; the PCWorld article itself was not read in full (headline and thread only), and someone else's Claude Code extension was not verified, it is a commenter's claim]
Firefox is now the last major browser that still supports uBlock Origin · HN thread, 601 points
topic 10Gergely Orosz on Grok Bot: "the OpenClaw moment for managed AI agents"
The last item is short, because the source limits it.
The Pulse of August 14, the only fresh Substack issue in the window. Orosz gives this the second slot, and in the free part the wording is: "Grok Bot: the 'OpenClaw moment' for managed AI agents? The Cursor team built and shipped a general AI harness that feels like 'the Codex experience, but for knowledge work'. I tried it, automated a lot of my daily workflows, and I'm impressed. More AI vendors will certainly copy this harness."
Then the paywall, and the rest was not read. So only what is on the page is reported: the issue also has a story about Meta failing to stop a resignation wave it triggered itself ("offering large retention stock packages, and it doesn't appear to be working"), but the details were not seen.
Following yesterday. Yesterday's misc quoted @awilkinson with the line "Grok Bot = Openclaw for normal people", and at the time it was one investor's joke. Today the same diagnosis comes independently from Orosz, the most technical source among the subscriptions, and from a practitioner who actually moved his workflows onto it. The claim got stronger in a day: from a quip it became an observation from two independent people.
Why it matters. Grok Bot does what homemade managed agents for knowledge work do, only as a vendor product. The argument for the homemade one is the same as in item 8: everything sits in git and can be read with your eyes. The argument against: a team stands behind the vendor harness, and one person and their evenings stand behind the homemade one.
No action follows from this. Swapping a live working setup for someone else's product because of two compliments in the feed is exactly the mistake yesterday's article about boring technology warned against. But one thing is worth putting on the radar: if the homemade version ever becomes the bottleneck, a market for such harnesses already exists. [fuzzy - only the free part of the issue was read, most of it is behind the paywall; Grok Bot itself was not seen, installed or tried; Orosz's enthusiasm is his personal experience with no measurements; the link to @awilkinson is two opinions coinciding, it is not confirmation]
The Pulse: Meta's self-inflicted resignation-wave (Grok Bot - second story) · @awilkinson yesterday: "Openclaw for normal people"
misc - briefly, what else is worth a look
• SWE Odyssey - a benchmark for the "ultra-long horizon" - @mattstallone: "As benchmark after benchmark saturates, we wanted a quantitative way to test the bigger claim: can agents work autonomously for hours and still build the right thing?" This is the same hole item 2 describes: benchmarks measure tasks that can be solved, and they do not measure work where you have to ask
• Anthropic published its second Risk Report (535K views) - @AnthropicAI: "Under the Responsible Scaling Policy we publish regular risk reports… The second risk report is now available." The report itself is a redacted PDF and was not analysed, so its contents are not summarized here: the link is given as is
• @garrytan: one person plus 20 agents against a whole department: "One person plus 20 agents running at the same time can outperform a whole engineering department at a Magnificent Seven tech company." Classic investor hyperbole, but note how it rhymes with item 2: more agents ≠ less need to ask
• @garrytan on Fable 5 at work: "the most amazing thing about using GStack before Fable 5 and after: now for many one-way-door questions Claude Code comes back with, you can just say 'take all recommendations' and be happy". This is the opposite assessment to item 2, which is why both are quoted
• @snowmaker: the size of YC's codebase over 14 years (431K views, 809 likes) - a chart from @snowmaker (Jared Friedman, YC). No commentary in the post itself, but 100+ comments in the thread: the classic "show a curve, let them argue"
• Toast 1 - a specialized model for search (189 points). The comment that explains the interest: "Deep sympathy for this idea of specialized LLMs for search. And it is extremely unclear how rough Google's entry here is"
• RISC-V: They should have known better (150 points) - a technical breakdown of RISC-V's architectural decisions. Nothing to do with AI, but it is the best engineering longread of the day
• Seven books I keep close because I love them (319 points) - Mark Dominus on seven books he keeps within reach. A rare case where the top of HN is simply a good text about reading
• @lennysan: "now a dental influencer" - a follow-up to yesterday's observation about AI in dentistry, which unexpectedly took off. An example of what lands best in the feed: an everyday observation
• @awilkinson: Things still has no API: "It's 2026, how does Things still not have an API?… not being able to interact with cloud AI is killing me." A small thing, but a precise marker of the new requirement for software: no API, no place in an agent workflow