jakozaur

I tested it, and it looks real but more nuanced. It is just for subagents; workflows do not seem affected. I guess many new users prefer workflows (ultracode), so the impact is less than the report suggested.

Moreover it is mostly cached reads, so subagents with short context cost little, while long-context ones cost much more.

In my Claude Code sessions, it looks more like ~8% at most.

Proxies are good if you pay API prices.

For subscriptions, it is easier to collect the JSONL files from ~/.claude/projects/ and similar directories. They are better organized (e.g., subagents) than what the proxy sees.

However, the best value comes from collecting agent sessions across the whole organization. You'll need to scrub credentials and handle security.

E.g., your prompt for simple auth works (e.g., decode JWT), but your sessions stop because you are flagged for cybersecurity risk.

The Fable is the worst offender, and it is not purely Anthropic's fault.

I worry we’ll have to approach this the other way around: verify that photos came from a camera, using hardware support like Apple’s Reference Image, rather than try to detect every AI generated one.

In many situations, photos are evidence. AI tools make convincing fakes, such as images of defect product.

A strong competitor in cybersecurity as an alternative to GLM-5.3 (Mistral reports 82% on CyberGym-E2E). Visual grounding is also impressive (42% on Dense 200 vs. 41% for GPT-6 Astra).

Otherwise, behind on the broader Pareto frontier, but not by much (Vals Index: 48.05% vs. GLM-5.3’s 53.51%; $13.78 vs. $7.25 per test). Many companies will prefer it over Chinease models.

Inflation in mentioned above example was 3.4% year over year for a professional at the higher end of the income scale. In that case it is inconvenience.

I agree that higher level of inflation might be brutal.

So many comments sound apocalyptic when they face inconveniences like inflation. See this recent post by a Polish guy:

https://news.ycombinator.com/item?id=49905487

Meanwhile, the Tigray war in Ethiopia may have killed as many as 600k, yet it barely registered in our public conversation. The same goes for Sudan.

Attention is a weird beast.

I'm Polish, but this is such a pessimistic, overdramatic take. The visualisations are impressive, but you're living in Poland's golden age. See this Hacker News story with 1,000+ upvotes: https://news.ycombinator.com/item?id=48062117

The war in Ukraine is a tragedy, but it has proved that Russia is weak. The price shock will speed up the transition to renewables, and Iran is losing its grip on Hormuz. So many of the other issues are solvable.

It's easy to be a pessimist, but there's a bright future ahead of us.

Typically, for AI training, you need human feedback. If AI were enough, then OpenAI would do it themselves. Though we all know how bad AI slop is, running it in a loop can lead to unreadable sentences.

They hired contractors on the condition that they provide human feedback without AI; those people broke the rules, so their contracts ended prematurely.

Not sure why this is news uncovered by an investigative journalist.

I don't believe we are anywhere close to AGI until I see 99.95% uptime at the frontier AI lab. Now they are undercounting. I hit 50x even when they claim they are up.

We now have alien intelligence that is very different from human intelligence.

I would say so. In this exercise, more tokens meant better results. Though DeepSeek tokens were as good as frontier ones.

In this hackathon, how much data you analyzed was correlated with tokens spent.

Analyzing the whole dataset gave the best insights, but cost more tokens. Using a small dataset could save tokens, but some insights were premature.

Yeah, real Jev got really weird, no benchmarking clause. Their Terms of Use (1(v)) and MCA (2.3(f)) both prohibit users from publishing "benchmarks or performance information about the Services". No major AI has it; we are back to Oracle-style legal.

Though Jev is original, it looks highly replicable.

I believe RTK would work well in that use case.

Sometimes creating less verbose variants yourself (a simple script, build.sh, with pointers to logs) can be a quick win.

The drop is not uniform; see Table I.2.6 (2022 -> 2025). Focusing on math:

Stable, above OECD average: Taiwan, Macao, Korea, Estonia, United Kingdom, Poland, Lithuania, Italy, New Zealand

Stable, around or below OECD average: United States, Slovakia, Greece, Croatia, Qatar, Montenegro, Mongolia, Mexico, Brunei

Dropping, previously above OECD average: Singapore, Japan, Hong Kong, Switzerland, Canada, Netherlands, Ireland, Australia, Belgium, Austria, Denmark, Finland, Germany, Sweden, Slovenia, Latvia, Portugal, Norway, Spain, France, Hungary

Dropping, around or below OECD average: Iceland, Israel, Malta, Serbia, Chile, Bulgaria, Kazakhstan, Argentina, Malaysia, Romania, Peru, Ukraine, North Macedonia

Improving, all below OECD average: Türkiye, Thailand, UAE, Philippines, Jordan, Georgia, Cambodia

Somewhat similar to the Claude Cowork pattern; Cowork is one toggle away from Chat.

Also follows the Grok Bot of an advanced computer use. OpenClaw for normies.

I tried it with a less typical case (the Warsaw, Poland transit app).

I'm impressed that it worked at all and got quite far. Though it is still broken in many ways (random geo locations), the menu bar on top...

The problem is that many web apps are quite simple, while native apps get trickier when it comes to UX. The LLM doesn't work well if it has to interpret animations or cross-correlate map data with locations on a map.

Two events: Anthropic did end up paying a $1.5B settlement for a case involving the use of pirated books. See Bartz v. Anthropic.

Second, apparently archiving content and providing it to others is risky business. It is often ruled as piracy, see Hachette v. Internet Archive. It went badly for Internet Archive, against common sense.

AWS rarely acquired startups, though leveraging someone else's tech is a typical pattern. It did license ParAccel and sold it as Redshift. It also runs Athena, which is based on Trino (a Presto fork).

Now, with Snowflake and Databricks earning big bucks in the intelligence era, time to gain market share by having DuckLabs under its ownership?

I wish AI leaders would fund more data archival projects, instead of hoarding it themselves. They could build a lot of goodwill, but the current state of the art is destroying old books and denying anyone access to them.

These archival projects have an enormous impact, though they are chronically underfunded. I wish there were a strong economic incentive to support them.

Yes, in the linked sources, SemiAnalysis did test for Codex too: https://x.com/SemiAnalysis_/status/2064815044085318040

Similar effect, but even 70 times higher compared to Claude, 40 times.

This isn’t limited to large system prompts. Coding-agent harnesses are also becoming more aggressive about using tools, even for trivial requests. In our tests, prompts such as “Hey” or “commit” sometimes triggered 30+ tool calls:

https://quesma.com/blog/the-true-cost-of-saying-hi-to-an-ai-...

Tokenflation seems very real: the number of tokens consumed by simple tasks keeps increasing.

Fable is also extended by another week till July 19: https://support.claude.com/en/articles/15424964-claude-fable...

It's rather hard to do at the proxy level with agentic coding, such as Claude Code or similar. These are long-chained sessions of tool use that heavily rely on prompt caching. Changing mid-flight is costly.

It looks like much more context is required to decide on the best model (e.g., summarizing logs might use a cheap model, whereas you likely want Opus/Mythos/GPT 5.6 to debug multithreading logic). In an agentic system, a decision about the model may be embedded in the decision to orchestrate the model.

Poland's first partly free election was on 4 June 1989, preceded by the roundtable negotiations.

The protests in Czechoslovakia came later, called the Velvet Revolution, from 17 to 28 November 1989. In June 1990, Czechoslovakia held its first democratic elections, a year after Poland.

Poland paved the way for the whole of central and eastern Europe. The Round Table produced the negotiated-exit template that Hungary built on in its own talks that summer, and that Czechoslovakia, East Germany, and the Baltics drew on as their regimes fell within months.

And it did so from the deepest macroeconomic crisis of any of the satellite states: hyperinflation running into the hundreds of percent by late 1989, an unresolved sovereign default from 1981, and chronic shortages.

Since then Poland has converged fastest of any of them. From a low base it has climbed to the upper-middle of central and eastern Europe by GDP per capita PPP, overtaken Hungary, and is now closing on Czechia and Slovenia.