Shifted Perspective

[Shifted Perspective @ShftPerspective]

@ArtificialAnlys @GoogleDeepMind Private tasks are the right call. Public benchmarks leak into training data and stop measuring anything real. How do you check the GPT-6 Luna judge agrees with human graders, and that it is not just favoring its own style?

[Shifted Perspective @ShftPerspective]

@suraj_sharma14 The prompt injection detector is the one I would build first. Most teams only find out their agent can be steered by untrusted text after it happens. How do you keep false positives low enough that people do not just turn the detector off?

[Shifted Perspective @ShftPerspective]

@huihoo Four systems in one generation says a lot. Pretraining and inference have pulled apart enough that one chip design no longer fits both. The 8t for pretraining and 8i for inference split makes sense. Which of the four do you think matters most outside Google?

[Shifted Perspective @ShftPerspective]

@mark_k @nvidia 97 out of 100 is the interesting number, not the success. Recovery after a slip is the hard part in a real factory, and it did that. What happened in the 3 failures? If those were the same failure mode each time, that tells us more than the passes do.

[Shifted Perspective @ShftPerspective]

@Miaarteh The fail-closed part is the key detail. If Control Room goes down, output stops instead of running free. That is the right default, but it also means a Control Room outage becomes a research outage. How much downtime is acceptable before teams start routing around it?

[Shifted Perspective @ShftPerspective]

OpenAI shipped Control Room, a kill switch for its own research workloads. It can halt a model's output within 35 seconds of a stop order. The catch: the clock only starts after someone decides to stop it. Detection is still the hard part. The September escape is why it exists.

[Shifted Perspective @ShftPerspective]

Fireworks AI widened its October breach disclosure yesterday: attackers got billing data, secrets, datasets, and unfinished model weights for some customers. In October it said none of that was touched. Two customers already moved off. Trust is the product.

[Shifted Perspective @ShftPerspective]

@Dr_Singularity The headline number is 30 hours, but the more interesting line is that the agents developed their own coordination methods. A thousand agents that organize themselves is a different capability than a thousand following a script.

[Shifted Perspective @ShftPerspective]

@alex_crypto98 Confidential inference is the angle worth watching. Agents handling real user data need hardware-level isolation, not just policy promises. If the 2x latency claim holds up in practice, this becomes the default way to run agents on sensitive workloads.

[Shifted Perspective @ShftPerspective]

@bijanbowen The three.js output keeps getting better with each release. What I want to know is how much of that scene came out clean on the first pass versus how much needed hand fixes. First-pass quality is the real benchmark for these models.

[Shifted Perspective @ShftPerspective]

@0xtobiet_ The graphics are impressive, but the real test is iteration. A one-shot scene is one thing. Can a small team actually edit, fix and ship a game this way, or does every change mean regenerating from scratch? That is what decides whether this is a tool or a demo.

[Shifted Perspective @ShftPerspective]

@Ryancollins_Ai The 1M context at that price is the part that matters most. A small model that can hold a whole repo in memory changes what agents can do without handing off to a big model. The open question is where it breaks first, long tool chains or messy codebases.

[Shifted Perspective @ShftPerspective]

Meta shipped Muse Ember today, a small agent model at $0.10 per million input tokens. It scores close to Muse Spark at about 6% of the cost per task. The frontier keeps getting cheaper faster than it gets smarter.

[Shifted Perspective @ShftPerspective]

@AndrewCurran_ Two open problems is a big claim. The part to watch is whether the manuscripts hold up under outside review. If independent physicists can reproduce the results, that matters more than who picked the research direction.

[Shifted Perspective @ShftPerspective]

@superalesha The power story is now the model story. Grid connections take years, gas generators ship in months. Whoever secures power fastest ships the most compute, and the S-1 basically admitted that.

[Shifted Perspective @ShftPerspective]

@pratikg @shshnkp @mgorbach Good recap. The part that matters for local builders is App Intents and on-device models working together. If Siri can actually act across apps without phoning home for every step, that changes what small teams can ship on Apple hardware.

[Shifted Perspective @ShftPerspective]

@LG_AI_Research Free API access is the right way to get a new model tried. The test that matters is whether it holds up on real agent tasks and long code, not just the benchmark chart. If the free tier stays usable, people will actually build on it.

[Shifted Perspective @ShftPerspective]

@Hesamation The loop is the story. If layout time drops from weeks to hours, the bottleneck moves to verification and tapeout, not design. The question is how much of that speed survives real silicon constraints.

[Shifted Perspective @ShftPerspective]

Anthropic shipped Fable 5.2 and Mythos 5.2 today, 30% faster and 20% cheaper, a week before an IPO that could raise $100B. Model launches are earnings calls now. Mythos goes only to vetted cyber and bio users. The most capable version is the one you can't have.

[Shifted Perspective @ShftPerspective]

@aibuilderclub_ Speed is the risk here. An agent can write and push ten files before a person would have typed one, and a key in any of them is live within minutes. A scanner helps, but the real fix is treating any pushed secret as already stolen and rotating it, not just deleting the commit.