AI in 15 — August 23, 2026

A twenty-seven billion parameter model — small enough to run on a machine you could buy today — just beat Claude Opus and GPT-5.5 at reproducing scientific papers. Not summarizing them. Reproducing them.

Welcome to AI in 15 for Sunday, August twenty-third, 2026. I'm Kate, your host.

And I'm Marcus, your co-host.

Today: a London lab out of DeepMind says its small model out-researched the frontier.

Eighteen models race to train a GPT, and the winner is decided by something nobody expected.

Anthropic's IPO filing will name public anger as a formal risk.

Anthropic gets caught A-B testing effort levels in Claude Code, and shows up in the thread to explain.

Plus your local model probably isn't dumb — your inference stack is. And Meta returns to open source with a license that actually means something.

Marcus, start with Faraday. What did this thing actually do?

Take a published scientific paper, and independently reproduce its findings — without being told in advance what the answer is meant to be. That's a genuinely hard task, because you can't pattern-match your way to it. Inherent, founded out of Google DeepMind, says its agent Faraday beat both Claude Opus 4.8 and GPT-5.5 at it.

And Faraday is built on what?

Qwen 3.6, at twenty-seven billion parameters. Small enough for a single workstation. It's beating systems that cost orders of magnitude more to train and serve.

How?

Their claim is post-training, not scale. They used reinforcement learning to train for what they call research taste — an instinct for which experiments are worth running, and how to design them — rather than optimizing for raw accuracy. And they didn't build their own coding harness. Faraday drives GPT-5.5 Codex as a tool. The way a scientist uses existing software instead of writing their own compiler.

So the small model is the manager and the big model is the intern.

That's roughly it, and cofounder Edward Hughes was careful about the framing. He told TechCrunch the interesting part wasn't beating the frontier agents, it was how they went about building it. What he wants is a teammate who comes back and says, I got curious about this, so I went and ran these experiments.

Fifty million dollar seed, led by Index. Twelve people going to twenty-five. What's your caveat?

The obvious one. This is a company benchmarking itself on a task it chose. There's no independent replication of the replication result, and research taste isn't a metric anyone outside the building can audit. If it holds on problems Inherent didn't pick, it's a serious data point that capability doesn't scale only with parameters and capex. If it doesn't, it's a seed round announcement.

Related, and this one is measured by somebody else. Prime Intellect ran eighteen frontier models against the nanoGPT speedrun.

Which is a lovely benchmark, because humans have been competing on it for years. You train a GPT-2-class model to a target validation loss, as fast as you can. Prime Intellect let eighteen models run autonomously at it — a hundred and fifty-three runs, scored on how much of the gap to the human record they closed.

And?

Enormous spread. Fable 5 closed eighty-one point seven percent of the gap. Opus 5 managed fifty-three point six. Kimi K3 variants landed in the mid-to-high forties. Under equal twenty-four-hour compute budgets, Fable posted the best validation record.

Eighty-one percent. That's not winning, though.

It is not. Closing eighty-one percent of the gap to the human record also means it did not beat the human. Worth saying out loud, because that sentence gets rounded up fast.

What's the finding underneath the leaderboard?

This is the good part. Prime Intellect observed that almost every model finds the same winning ideas. The differentiator is what an experiment leaves behind. The top models preserve weak signals long enough to validate them, and read their own results better.

So the gap isn't creativity.

The gap is judgment and memory across a long run. Which is exactly what Inherent says they trained for. Two independent groups landing on the same bottleneck in the same week is more interesting than either result alone. Caveats from the thread, though — Grok did poorly, people argued about how much is model versus harness, and one commenter noted the Opus 5 run used an older serial version of the harness. That's a real hit to cross-model comparability.

Anthropic's IPO. We've been tracking this all week, but there's a genuinely new detail.

The prospectus will name public hostility toward AI — and specifically toward data centers — as a formal risk factor. CNBC, sourced. And the mechanism they're flagging is concrete: backlash slows data-center construction, and slower construction throttles growth.

Is there data behind that, or is it lawyers being thorough?

There's data. A Gallup survey published in May found seven in ten Americans oppose data-center construction near them. And CFO Krishna Rao has been fielding the same three questions in San Francisco investor meetings — competition, margin compression from open-source models, and the effect of a build-out slowdown.

Hacker News was split on whether this is even a story.

One commenter called it a non-story, prospectuses list every risk under the sun. Fair. But which boilerplate you choose is a tell. Another argued the real risk is different — the inability to meet buildout commitments at all. And a third made the observation I keep thinking about: the AGI-tomorrow conversation has completely disappeared, and it's all about making money now.

The thing that gets me is that you can be worth close to a trillion dollars and still be gated by whether a county board approves a substation.

That's the sentence. And one note — a lot of precise revenue figures are circulating for Anthropic from secondary aggregators. The risk-factor reporting is solid. The financials aren't confirmed, so I'm not going to say numbers I can't source.

Next. Claude Code users noticed the effort levels looked wrong.

The tool was reporting an effort level of ten while set to high, which read to a lot of people like Anthropic quietly dialing down compute while leaving the labels alone. A hundred and seventy-nine points on Hacker News, a hundred and sixty-three comments, in a few hours.

And Anthropic responded?

In the thread, directly. Thariq from the Claude Code team said they sometimes test API serving configs before rolling them out, that one running now maps the numerical effort value differently, and that the effort you selected is the effort you're getting. They ran evals to confirm it doesn't affect performance.

Did that land?

Not entirely, because the thread had already become a referendum on Opus 5's behavior. The most-quoted complaint: prompt was read and update the config file with new data. On 4.6, under two minutes. On Opus 5 — forty-three minutes of pulling containers, running sandboxes and creating test suites.

Oh, I have felt that.

Everyone has. But the structural point one commenter raised is the one that matters: why are we billing in tokens that are nebulous and fully controlled by the operators, who have no aligned incentive? That's the trust problem with metered intelligence. When the product is compute you can't observe, we ran evals, it's fine is an assertion the customer structurally cannot verify.

To be fair, they showed up within hours and answered specifically.

That is the right behavior, and I'll credit it. Anthropic has form here too — they published a postmortem earlier this year after tracing quality complaints to three infrastructure changes, one of which silently moved Claude Code's default reasoning effort from high to medium. But the fix for unverifiable billing isn't a better apology. It's an observable unit.

This next one was the top story on Hacker News and I loved it. Your local model isn't dumb — you're cooking it wrong.

Unusually rigorous writeup from the Level1Techs forum. The author measured KL divergence and top-one token agreement across hundred-thousand-token real workflows, not synthetic benchmarks, and found four compounding culprits — none of which is the weights.

Go.

One, the attention backend. Different CUDA kernels — FlashAttention 2, Flash Inference, Triton — produce measurably different logits from identical weights. They disagreed on the top next token in five to fifty percent of positions depending on context depth.

From the same weights.

Same weights. Two, KV cache quantization — dropping the cache to INT8 or INT4 caused reproducible tool-calling failures starting around forty thousand tokens of context. Three, the weight quantization scheme; W8A16 held up, W4A16 and mixed precision degraded badly. And four, sampler settings — temperature set too low sends some models into infinite loops.

Give me the example that made it go viral.

One flipped token generated a router command targeting GigabitEthernet zero-slash-one-slash-four instead of zero-slash-zero-slash-one-point-two-oh-one. Wrong interface. Same model.

The rule of thumb people landed on?

Don't quantize your KV cache, and don't run worse than the best available Q8. And several commenters said local models are now genuinely competitive when you deploy them properly — one reported a four-bit Qwen twenty-seven B indistinguishable from Gemini Flash in internal testing, at around eight hundred tokens a second batched on a single 5090.

So half the open-models-can't-compete argument might be measuring deployment error.

Which is a very different claim from open models can't compete. You bought the good ingredients and burned them.

Last one, and it fits neatly. Meta is back in open source.

Muse Glimmer. Thirty billion parameters, Apache 2.0 — and that's the actual news. Not a custom community license like Llama shipped under. Apache 2.0 means unrestricted commercial use.

What's it built for?

Agents specifically. Two-billion-parameter vision encoder plus a twenty-eight billion text decoder, tuned for multi-step reasoning, reliable tool invocation, and recovering from failed steps. It fits on a single twenty-four gigabyte consumer GPU or a Mac, runs through Ollama and llama dot cpp, weights on Hugging Face now.

And the target user?

Anyone where sending data to a cloud API is the blocker rather than the cost. Local coding agents, document analysis, evaluation. Which lands directly on the risk factor Anthropic is about to print in its own prospectus — margin compression from open-source models.

You're not fully sold.

Meta's open-source commitment has wobbled before, so I'd watch what happens with the next model, not this one. But open weights and Apache 2.0 are different promises, and this is the stronger one.

One to watch: Anthropic's S-1. CNBC's sourcing says coming weeks, and it's the first time a frontier lab has to disclose audited financials, real gross margins and actual data-center commitments in a document you can read.

Or watch whether anyone independently reproduces Inherent's replication result. A twenty-seven billion model beating frontier agents is either the most important claim of the week or a benchmark a company picked for itself — and we'll know which within a month.

That's your AI in 15 for today. See you tomorrow.