Daily AI News · Nov 6 Midday Edition: Nasdaq Plunge and Evals Pricing

Another busy day in AI — we've got stuff from the markets, from leaderboards, and from the developer community.

Today we're covering a few things: first, Vercel's Token Pass unified billing; second, a job offer from the evals leaderboard world; third, the Nasdaq plunge and AI worries; fourth, a Codex tutorial; fifth, Anthropic's issue backlog. Plus deep dives on deepseek^Ad optimization, agent tutorials, and multi-agent bug hunting.

Yeah.

Alright, let's start with the news.

Just saw an update from Vercel — Token Pass credits no longer need a dedicated key, they just automatically apply to AI Gateway calls now.

Yeah, that's actually pretty practical. Before, credits were tied to a dedicated key, so you had to remember which key to use. Now standard keys and OIDC both get the credits.

So does it cut off when you run out? That's what I care about.

Requests from standard keys and OIDC automatically fall back to the team's AI Gateway credits — service keeps running. But if you're still on a Token Pass dedicated key, when credits run out it just stops until the next cycle.

So anyone still on the dedicated key needs to switch credentials fast, otherwise they miss out on the fallback.

Exactly. And privacy isn't relaxed either — Token Pass requests still get zero data retention, and fallback follows the team's ZDR settings.

Second item — something fun on Twitter. Someone at OpenAI praised a particular evals benchmark approach, and then publicly offered Dwarkesh Patel a job designing evals.

I think that signal is really interesting. The gap in general model capability is shrinking, so the quality of eval design now directly determines how people judge models and where training goes.

So a sloppy benchmark basically rewards gaming the score, not real ability.

Exactly. How tasks are constructed, how scoring criteria are set, how data contamination is controlled, how adversarial samples are designed — all of that needs dedicated judgment.

And what they saw in Dwarkesh was that sensitivity to good questions from his interviews — that transfers to an evals role.

Next, a big one. Hacker News is circulating a Reuters report — the Nasdaq Composite's selloff this round has wiped out roughly 2.5 trillion dollars in market value, and the market blames it on a political shift called the Blue Wave.

2.5 trillion — that number is definitely scary on its face. But look at what's falling — it's the policy and capital assumptions built around AI.

You mean the tech trend itself isn't being rejected?

Right. Investors are worried the new policy cycle might tighten AI regulation, touch taxes and antitrust — that changes big tech's commercialization pace and return expectations.

No wonder the panic amplifies — Nasdaq's weight is all concentrated in the main beneficiaries of the AI narrative, so it all moves together.

So it's the details of regulation and tax policy landing that will determine whether this is short-term sentiment or a long-term valuation reset.

Then there's one from Bilibili — a hands-on Codex tutorial, from install and login all the way to Skills and hooking up external tools via MCP.

There's huge demand for that kind of tutorial now — people don't want concepts, they want to know how to actually plug it into their workflow.

Yeah, especially the MCP part — how to connect external tools, that's where a lot of people get stuck.

Last news item — someone on Reddit asked Anthropic directly: you do 300,000 code merges a day, but Claude Code's public repo has over 16,000 open issues, more than 98% unclaimed, including some serious security bugs.

That contrast is definitely striking. But you have to admit — internal merges and open-source issue handling aren't the same thing.

But the author's point is that this ratio reflects selective neglect in process design, not just a lack of people.

He also pointed to a User-Agent privacy issue sitting for months, saying they can build pixel heart easter eggs but can't use Claude to triage issues. The technical bar is low — it's governance and priorities that block it.

Alright, that's the news. Now let's go deep on two things — the one saying deepseek can still optimize 27x on tokens, and the multi-agent bug-hunting comparison.

Sure, let's start with the optimization one — I'm curious how they calculated it.

There's a video on Bilibili with a catchy title — your deepseek can actually optimize 27x on tokens, stop eating empty calories. Over 380,000 engagement score, pretty hot.

Set the 27x number aside for a second — I think the point is token utilization. In a lot of calls, the prompt and context have tons of repetition and redundancy.

So it's like dieting — what you eat doesn't turn into energy, it all becomes empty calories.

That metaphor is actually pretty accurate. Think about it — for the same task, the messier the context, the more likely the model drifts, and the more you pay.

But 27x optimization sounds a bit exaggerated, right? Is that really doable?

Depends on the baseline. If it's highly repetitive batch calls, like templated tasks, compressing context and reusing cache really can save a lot. But in general scenarios it's hard to get that big a multiple.

So that number is more like an upper bound for an extreme case, not everyone's everyday gain.

Right, and that's not the main point — the main point is it reminds people not to treat tokens like free water. Context management itself is a product capability.

Then the flip side — if you cut context too aggressively to save tokens, do you cut out necessary info too?

Yes, that's the balance point. You cut redundancy, not intent. Judging what's redundant relies on understanding the task, not a one-size-fits-all approach.

So the conclusion is, 27x is an eye-catching claim, but what really works is strategic context compression and caching — that's the right path to saving tokens.

Agreed. And that kind of optimization matters especially for agent products, since agents make way more calls to begin with.

Speaking of agents, there's also a full beginner tutorial on Bilibili — MCP plus Agent Skills, multi-agent, workflow automation, project practice, all in one.

These tutorials are popular for a reason — people have moved from asking what an agent is to asking how to build one that actually does work.

But a full beginner course like that — I'm a bit skeptical. Can you really go from zero to a working agent with one tutorial?

A tutorial gives you the skeleton, but a working agent — the hard part is tool definition and error handling, and you have to grind those out in your own scenario.

So that's actually the same story as the token-saving topic — more agent calls means context and cost both pile up.

Right, connect the two and it's clear. Whether agents scale is half model capability, half engineering — controlling cost and context.

Last one, and I'm especially interested in this. There's a YouTube video where one AI found 269 bugs, while two coding agents found 86 and 59 respectively.

That comparison is interesting. One AI found 269, two agents together only 145 — that's not a small gap.

So that's the opposite of the multi-agent collaboration hype, right? Isn't multi-agent supposed to be stronger?

Hold on — first look at how they define a bug. If the single AI is casting a wide net, of course the count is higher. Agents might be more focused but not necessarily as comprehensive.

So these numbers don't necessarily show agents are weak — they show the evaluation criteria are different.

Right. And coding agents have tool calls and execution environments, so the bugs they find might lean more toward reproducible ones, while a single AI leans more toward static scanning.

So we're back to the evals topic — how the benchmark is designed directly determines who looks stronger.

Exactly. So when I connect these three — eval design, cost optimization, agent collaboration — they're all saying the same thing: beyond capability, engineering and evaluation systems are what decide who wins.

To sum up — don't get carried away by a single number; understanding the criteria and the scenario matters way more than chasing multiples.

Yeah, this episode has a lot in it, but the core is just that one line.

Alright, that's it for today — pretty dense, take your time digesting. See you tonight!