Chai Discovery Raises $400M Series C; AI-Designed Antibodies Are Already Heading to Big Pharma
AI drug discovery startup Chai Discovery has closed a $400 million Series C round. The headline here is less the amount than the fact that AI-designed antibodies are already being folded into major pharmaceutical companies' pipelines. A few years ago, "an AI produced a drug candidate" was news in itself; now we're at the stage where it connects directly into big pharma's commercial processes.
What matters in the context of this raise is that the yardstick for drug discovery AI has shifted from "good scores in demos" to "does it hold up as an actual clinical trial candidate." You could say we've entered a phase where the question is no longer just model performance, but how well it integrates with in-house chemistry and bioassay expertise.
Simon Willison Revisits the "AI Lethal Trifecta"
Simon Willison, who coined the term "prompt injection," has laid out the three conditions that spell disaster for AI agents as the "Lethal Trifecta."
Access to private data (internal databases, email, etc.)
Exposure to untrusted external content (web pages, external text, etc.)
A channel for exfiltrating data (tool calls, outbound communication)
When all three are present, malicious content can hijack the agent and open a path for internal data to be sent outside. Willison has been pointing this out for a while, but it's drawing renewed attention on Hacker News because people are starting to internalize how, as agent deployments become commonplace, every "feature add" is also a risk add.
A "Gandalf-style puzzle" has also been released for hands-on learning. It's structured as ten levels where you actually attempt injection at each step, with the final level challenging you to line up all three conditions for a breakthrough — a beginner-friendly teaching resource. Agent designers would do well to try it themselves and check which of the conditions their own designs satisfy.
Three Local LLM Stories: 1M Context, 120B on a Phone, and a Coding Comparison
Over in the LocalLLaMA world this week, we saw a run of benchmarks close to real-world usage.
**1. DeepSeek V4 Flash on RTX 5090 + llama.cpp**
There's a report of someone running a 1M (roughly one million token) context using the Unsloth build of `DeepSeek-V4-Flash-UD-Q8_K_XL` with recent `llama.cpp` changes. Prefill hits 650–700 tok/s, decode runs at 17 tok/s, and loading takes 32 seconds. Not as fast as the Qwen family, but reportedly entering practical territory.
Another user compared IQ3_XXS-AS and IQ2_S on an RTX 3090 + 128GB DDR4 setup, testing both mainline llama.cpp b10064 and fairydreaming's dsv4 fork. Their conclusion: "the fork is no longer needed" — optimizations merged into mainline have all but erased the gap. That should make maintenance cheaper for individual operators too.
**2. Running GPT-OSS-120B on an Android Phone (CPU Only)**
A case of running the 60GB gpt-oss-120b Q4_K_M on a OnePlus 15R (about 11GB of usable RAM) has been making the rounds. A plain mmap load gets 0.089 tok/s, but by exploiting the MoE architecture to stream only the needed experts from flash, that's pushed up to 1.3 tok/s. That works out to roughly a 14x efficiency gain from streaming, even with the model at 5x the size of memory.
Interestingly, the output is identical to a fully RAM-resident setup, with no GPU or NPU involved. It's a clever use of the MoE-specific property that only a subset of experts is used per token — a useful design reference for edge inference.
**3. Gemma4-31b vs Qwen3.6-27b: A Coding Comparison**
On a more everyday note: a user who had been running Qwen3.6-27b as their main coding agent switched to Gemma4-31b and found the back-and-forth over bugs dropped, letting their prototype leap forward. That said, they've settled on a setup of "Gemma as the coder, Qwen as the reviewer/QA," which lets Qwen's strengths shine — an interesting case showing that dividing roles beats swapping models, with a real feel for how agent design plays out in practice. It's one person's impression, so don't over-generalize, but it's a useful reference when designing workflows where multiple models collaborate.
AWS Rolls Out Amazon Quick, an Agent for Sales Teams
AWS has started deploying "Amazon Quick" as an AI assistant for sales organizations. 3M and AWS Global Sales are already on board, offloading the "non-selling time sinks" — CRM updates, prospect research, email drafting — to Quick.
Integrations with existing tools like Microsoft 365 and Outlook are also on offer, making the product design clear: position the agent as "a colleague for a specific job." A mobile layout for the Quick dashboard has also been added, so KPIs and charts can be scrolled in a single column from a phone. These small improvements, clearly designed around real field usage, are a useful gauge of how deeply agents are penetrating everyday work.
Academic Tool Prism Compiles and Returns Someone Else's Paper
A bug has been reported in Prism, a tool used among researchers, where the compile process returned another user's paper. After the issue was raised on Discord, the site was taken private within 10 minutes — but users are now worried about "whether my paper has leaked somewhere."
Precisely because input confidentiality is critical for academic AI tools, we're waiting to hear the root cause analysis and prevention measures. With several similar tool incidents reported recently, this is another reminder that AI use in research means weighing "convenience" against "data leakage risk."
The Atlantic Takes Another Look at AI's Water Use
The Atlantic has published an investigation into data center water consumption. When we talk about AI's environmental footprint, electricity tends to get the attention — but cooling water consumption can also have serious impacts depending on the region. As data centers keep being built and expanded in arid regions, how to live alongside local water resources is being questioned anew.
Regulators are getting active too: some municipalities have begun pivoting to construction freezes or conditional permits for data centers. Alongside "decarbonization of electricity," "water sustainability" is emerging as a new constraint on data center siting — a point worth keeping on your radar.
Wrap-up
What emerges from this week's news is a sense that AI has shifted from "feature demo" to "on-the-ground infrastructure and decision material." From drug discovery pipelines to sales workflows, research support, and every corner of local inference tinkering, AI is becoming a built-in assumption of the work. In turn, how we handle "trust and side effects" — safety (the Lethal Trifecta), data leakage (Prism), environmental impact (water use) — has moved to the forefront as the next design challenge.