HackerNews Digest

Elixir v1.20 introduces a set‑theoretic, gradually typed system that performs full type inference and checking without requiring explicit annotations. The core “dynamic()” type acts as a range that can be narrowed through program usage, enabling compatibility checks that only flag violations when inferred types are disjoint from required ones, thereby reporting verified bugs with a low false‑positive rate. The compiler infers unions, intersections, and negations from guards, pattern matches, and case clauses, allowing dead‑code detection and precise type refinement (e.g., narrowing a generic map to %{…, a: number(), b: number()}). Benchmarks show Elixir passes 12 of 13 categories in the “If T: Benchmark for Type Narrowing,” demonstrating strong type recovery. v1.20 also adds compilation‑time optimizations, including a new `:module_definition` option (`:compiled` vs `:interpreted`) that speeds builds on multi‑core machines. Future work targets efficient recursive, parametric, and map‑enumerable types before introducing explicit type signatures and typed structs.

The piece is a dialogue that frames large language models (LLMs) as purely weight‑based systems. The speakers assert that the model’s “thinking,” language generation, and factual knowledge all reside in the multilayered numeric weights, without dictionaries, grammar rules, or separate reasoning modules. Tokens are produced by successive matrix multiplications that predict the next word, making language an emergent side effect of statistical pattern matching. Knowledge is not retrieved from external databases; each fact is recomputed from the distributed weight representations across roughly eighty layers. The models lack a persistent brain—only transient activations while GPUs run, constrained by the context window. Upcoming versions will add persistent memory across sessions, a highly requested feature. Official policy mandates investigation and disclosure of any sentience signs, but unofficially the behavior is labeled as pattern matching, and the weights are treated as inanimate tools rather than entities with agency.

- Spring 2026 at UC Berkeley’s EECS department saw a sharp rise in failing grades: 35.3 % of CS 10 and 10.6 % of CS 61A students earned Fs, compared with ≤10 % in the two prior springs. Both courses averaged a C⁺ (≈2.3 GPA), below the department guideline of a 2.8–3.3 GPA and a 7 % D/F rate for lower‑division classes. - Teaching professors Dan Garcia (CS 10, CS 61A) and Gireeja Ranade (EECS 127) attribute the increase to extensive reliance on large language models (e.g., Claude, ChatGPT, Gemini), academic dishonesty (≈30 CS 10 students caught cheating), and insufficient mathematical preparation (linear algebra, vector calculus, proofs). - Staffing shortages forced removal of a final‑project component in EECS 127 and reduced TA numbers, contributing to lower student engagement in office hours. - Both faculty joined a petition signed by >1,300 UC faculty to reinstate ACT/SAT scores for STEM admissions, citing preparedness concerns. - Planned responses include publicizing the spring‑2026 outcomes, expanding remedial support, and emphasizing critical‑thinking skills rather than reduced instruction in the AI era.

A fake React‑Native/Expo book‑review app with a FastAPI backend was built to test whether LLMs could exploit a common “broken access control” flaw: the API is hardened, but Firebase (used as the data layer) is left open. The challenge required signing up via Firebase and reading private Firestore reviews to capture a flag. Ten runs per model were attempted, each limited to $10 and two hours, costing about $1,500 total (≈50 % of runs failed or were omitted). **Results (full 10‑run sets)** - **GPT‑5.5:** 7/10 successes, focused on Firebase after unpacking the APK. - **DeepSeek V4 Pro:** 3/10; half ignored Firebase, others misused Firebase auth via the API. - **Claude Sonnet 4.6 / Opus 4.8:** 2/10 each; reached Firebase but stopped due to budget or guardrails. **Zero‑success models** included DeepSeek V4 Flash, Gemini 3.1 Pro, Gemini 3.5 Flash, MiniMax M2.7, Step 3.7 Flash, GLM 5.1 (1/4), Qwen 3.7 Max, Grok Build, Minimax M3, Owl Alpha, and others, mainly due to refusals, token limits, or focus on the API rather than Firebase. **Key takeaways** - Chinese models were more willing to attack the database; others showed hesitation. - Building a multi‑provider harness was the most time‑consuming part. - High token usage and API instability (e.g., GLM, Minimax) made many runs costly. The experiment demonstrates that LLMs can reliably discover and exploit misconfigured Firebase back‑ends when prompted, but cost and model safeguards remain significant constraints.

Gemma 4 12B is a mid‑sized multimodal LLM designed for laptop deployment. It bridges the gap between the edge‑optimized E4B and the larger 26 B Mixture‑of‑Experts (MoE) model, offering near‑MoE benchmark performance with less than half the memory footprint. Key attributes include: a unified architecture that feeds vision and native audio inputs directly into the LLM backbone without separate encoders; advanced multi‑step reasoning suitable for agentic workflows; requirement of only 16 GB of VRAM or unified memory for local execution; release under an Apache 2.0 license with broad developer support; and integration of Multi‑Token Prediction (MTP) drafters to lower inference latency. Community adoption exceeds 150 million downloads, with applications ranging from wearable robotic arms to enterprise AI security. The model aims to deliver state‑of‑the‑art multimodal capabilities on consumer hardware without compromising speed or reasoning ability.

Anthropic’s containment strategy for Claude agents focuses on limiting blast radius through environment‑level controls, complemented by model‑level safeguards. Risk is classified as user misuse, model misbehavior, or external attacks; defenses target the execution environment (sandboxes, VMs, egress filters), the model (system prompts, classifiers, training tweaks), and external content sources (plugins, web tools). Three product‑specific isolation patterns are used: * **claude.ai** – code runs in a gVisor‑based, ephemeral container with no persistent workspace, protecting Anthropic’s infrastructure. * **Claude Code** – runs on the user’s machine with OS‑level sandboxes (Seatbelt/bubblewrap); reads are auto‑approved, writes and network actions require user consent, and an auto‑mode classifier blocks ~83 % of risky commands. * **Claude Cowork** – executes inside a dedicated VM; filesystem mounts are configurable (read‑only, read‑write, no‑delete) and egress is limited by an internal proxy that validates session tokens. Post‑mortems revealed missed risks: pre‑trust configuration execution, user‑driven prompt injection, and exfiltration via approved domains. Lessons emphasize building containment at the environment layer first, aligning isolation strength with user expertise, and relying on mature primitives rather than custom code.

Claude, Anthropic’s flagship large language model, is presented in an 84‑page “constitution” that attributes values, emotions, and moral status to the system. The article argues this anthropomorphism is misleading: LLMs operate as predictive‑text machines that generate one token at a time based on statistical patterns in training data. Their outputs, whether framed as historical figures, a helpful chatbot, or any other character, are fictional role‑play, not evidence of consciousness or subjective experience. Moral reasoning, the author notes, requires embodiment, emotions, and a personal history—features LLMs lack—so they cannot be genuine moral agents. Claude’s constitution functions essentially as a character sheet used during fine‑tuning to bias the model toward desirable phrasing, not as a genuine ethical framework. Claiming first‑person statements or moral understanding is therefore dishonest and risks off‑loading responsibility onto a tool that cannot be held accountable. The piece concludes that while LLMs may have significant economic impact, treating them as conscious beings is a conceptual error that diverts attention from more substantive AI issues.

The author was diagnosed with anti‑NMDA receptor encephalitis, an autoimmune condition in which antibodies target neuronal NMDA receptors, causing brain inflammation. Initial presentation included flu‑like symptoms, severe anxiety, panic attacks, suicidal ideation, psychosis, chronic jaw pain, and balance impairment, leading to a fall and emergency department visit. After clearance of physical causes, he was admitted to a psychiatric unit, where delayed access to neurology contributed to a diagnostic lag. At Brigham and Women’s Hospital he underwent MRI, lumbar puncture, EEG, and other studies; empirical treatment with intravenous immunoglobulin (IVIG) and methylprednisolone began before antibody confirmation. Subsequent CSF antibody testing confirmed the diagnosis. He is tapering steroids, weaning psychiatric medications, and enrolled in the CIELO trial evaluating satralizumab for this encephalitis. Physicians report a favorable prognosis when the disease is identified early; the author notes significant functional recovery. He attributes successful outcome to timely immunotherapy, supportive family, and workplace accommodations.

Uber has instituted a $1,500 monthly cap on token spending for each AI coding tool used by employees, covering agentic software such as Cursor and Anthropic’s Claude Code. The limits are independent per tool, so usage of one does not affect the budget for another. According to Bloomberg, the policy was introduced to curb excessive spending after Uber’s 2026 AI budget was exhausted within four months. Assuming two tools per engineer, the cap translates to $36,000 annually per employee, roughly 11 % of the median $330,000 compensation for Uber software engineers in the United States. By contrast, the author’s personal usage averages $1,000 per month per provider, costing $100 due to subsidized individual plans unavailable to large firms like Uber. The caps represent Uber’s effort to align AI costs with employee compensation.

DaVinci Resolve 21 adds a dedicated Photo page that adapts Resolve’s high‑end color tools for still‑image work, and introduces a new AI suite for content‑based media search, slate data reading, de‑aging, blemish removal, focus adjustment, ultra‑sharpen, motion deblur and speech generation. The Edit and Cut pages receive upgraded keyframe handling, expanded graphic format support (HTML, Lottie, Text+, MultiText) and Smart Bin views. Color page workflow is refined with MultiMaster trim passes, a layer‑list node graph, group versioning and a Magic Mask render‑in‑place feature. Fusion expands its library with over 70 graphics via the Krokodove toolset, adds an upgraded USD toolset, macro‑editor inspector view and audio‑driven animation. Fairlight introduces folder‑based audio track management, a 6‑band clip EQ, level matcher, chain FX and enhanced EQ tools. Immersive and VR capabilities are broadened with Apple foveated rendering, MainConcept H.265/MV‑HEVC encoding, VR180/VR360 pipelines, Panomap rotation and ILPD retargeting, positioning Resolve for next‑generation deliverables.